program(1.3) [buildInfo = dict({{"coremlc-component-MIL", "3405.2.1"}, {"coremlc-version", "3405.2.1"}})] { func main(tensor x) [FlexibleShapeInformation = tuple>>, tuple>>>>((("DefaultShapes", {{"x", [12, 64, 1, 2]}}), ("EnumeratedShapes", {{"25a592c0", {{"x", [28, 64, 1, 2]}}}, {"38088a33", {{"x", [18, 64, 1, 2]}}}, {"42bba9e4", {{"x", [22, 64, 1, 2]}}}, {"59316af0", {{"x", [12, 64, 1, 2]}}}, {"6bca635d", {{"x", [16, 64, 1, 2]}}}, {"84aa3ba0", {{"x", [8, 64, 1, 2]}}}, {"981a1dc8", {{"x", [32, 64, 1, 2]}}}, {"9d44cf0c", {{"x", [24, 64, 1, 2]}}}, {"9e64e8ea", {{"x", [10, 64, 1, 2]}}}, {"bcc527a0", {{"x", [26, 64, 1, 2]}}}, {"ce7d4a38", {{"x", [20, 64, 1, 2]}}}, {"d8a49046", {{"x", [14, 64, 1, 2]}}}, {"efb221f6", {{"x", [30, 64, 1, 2]}}}, {"f0e65740", {{"x", [1, 64, 1, 2]}}}})))] { string x_3_pad_type_0 = const()[name = string("x_3_pad_type_0"), val = string("valid")]; tensor x_3_strides_0 = const()[name = string("x_3_strides_0"), val = tensor([1, 1])]; tensor x_3_pad_0 = const()[name = string("x_3_pad_0"), val = tensor([0, 0, 0, 0])]; tensor x_3_dilations_0 = const()[name = string("x_3_dilations_0"), val = tensor([1, 1])]; int32 x_3_groups_0 = const()[name = string("x_3_groups_0"), val = int32(1)]; tensor var_53_to_fp16 = const()[name = string("op_53_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; tensor layer_in_proj_bias_to_fp16 = const()[name = string("layer_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131200)))]; tensor x_3_cast_fp16 = conv(bias = layer_in_proj_bias_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_53_to_fp16, x = x)[name = string("x_3_cast_fp16")]; fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)]; tensor fill_like_0_cast_fp16 = fill_like(ref_tensor = x_3_cast_fp16, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")]; tensor var_70_begin_0 = const()[name = string("op_70_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_70_end_0 = const()[name = string("op_70_end_0"), val = tensor([0, 1024, 1, 1])]; tensor var_70_end_mask_0 = const()[name = string("op_70_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_70_cast_fp16 = slice_by_index(begin = var_70_begin_0, end = var_70_end_0, end_mask = var_70_end_mask_0, x = fill_like_0_cast_fp16)[name = string("op_70_cast_fp16")]; tensor var_71_to_fp16 = const()[name = string("op_71_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133312)))]; tensor special_tokens_cast_fp16 = mul(x = var_70_cast_fp16, y = var_71_to_fp16)[name = string("special_tokens_cast_fp16")]; int32 var_74 = const()[name = string("op_74"), val = int32(3)]; bool x_5_interleave_0 = const()[name = string("x_5_interleave_0"), val = bool(false)]; tensor x_5_cast_fp16 = concat(axis = var_74, interleave = x_5_interleave_0, values = (special_tokens_cast_fp16, x_3_cast_fp16))[name = string("x_5_cast_fp16")]; int32 var_86 = const()[name = string("op_86"), val = int32(-2)]; int32 var_90 = const()[name = string("op_90"), val = int32(1)]; int32 var_95 = const()[name = string("op_95"), val = int32(2)]; fp16 const_1_promoted_to_fp16 = const()[name = string("const_1_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_100_cast_fp16 = mul(x = x_5_cast_fp16, y = const_1_promoted_to_fp16)[name = string("op_100_cast_fp16")]; bool x_7_interleave_0 = const()[name = string("x_7_interleave_0"), val = bool(false)]; tensor x_7_cast_fp16 = concat(axis = var_90, interleave = x_7_interleave_0, values = (x_5_cast_fp16, var_100_cast_fp16))[name = string("x_7_cast_fp16")]; tensor out_1_axes_0 = const()[name = string("out_1_axes_0"), val = tensor([1])]; fp16 var_110_to_fp16 = const()[name = string("op_110_to_fp16"), val = fp16(0x1.5p-17)]; tensor out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_110_to_fp16, x = x_7_cast_fp16)[name = string("out_1_cast_fp16")]; tensor layer_encoder_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_0_input_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135424)))]; tensor out_3_cast_fp16 = mul(x = out_1_cast_fp16, y = layer_encoder_layers_0_input_layernorm_weight_to_fp16)[name = string("out_3_cast_fp16")]; tensor var_116_split_sizes_0 = const()[name = string("op_116_split_sizes_0"), val = tensor([1024, 1024])]; int32 var_116_axis_0 = const()[name = string("op_116_axis_0"), val = int32(1)]; tensor var_116_cast_fp16_0, tensor var_116_cast_fp16_1 = split(axis = var_116_axis_0, split_sizes = var_116_split_sizes_0, x = out_3_cast_fp16)[name = string("op_116_cast_fp16")]; string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")]; tensor query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor([1, 1])]; tensor query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor([0, 0, 0, 0])]; tensor query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor([1, 1])]; int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)]; tensor var_81_to_fp16 = const()[name = string("op_81_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139584)))]; tensor 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_81_to_fp16, x = var_116_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 key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor([1, 1])]; tensor key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor([0, 0, 0, 0])]; tensor key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor([1, 1])]; int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)]; tensor var_82_to_fp16 = const()[name = string("op_82_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2236800)))]; tensor 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_82_to_fp16, x = var_116_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 value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor([1, 1])]; tensor value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor([0, 0, 0, 0])]; tensor value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor([1, 1])]; int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)]; tensor var_83_to_fp16 = const()[name = string("op_83_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2499008)))]; tensor 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_83_to_fp16, x = var_116_cast_fp16_0)[name = string("value_states_1_cast_fp16")]; tensor concat_0x = const()[name = string("concat_0x"), val = tensor([-1, 16, 64, 3])]; tensor embed_1_cast_fp16 = reshape(shape = concat_0x, x = query_states_1_cast_fp16)[name = string("embed_1_cast_fp16")]; tensor concat_1x = const()[name = string("concat_1x"), val = tensor([-1, 2, 64, 3])]; tensor embed_3_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("embed_3_cast_fp16")]; tensor concat_2x = const()[name = string("concat_2x"), val = tensor([-1, 2, 64, 3])]; tensor value_states_3_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("value_states_3_cast_fp16")]; tensor cos_to_fp16 = const()[name = string("cos_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2761216)))]; tensor var_142_cast_fp16 = mul(x = embed_1_cast_fp16, y = cos_to_fp16)[name = string("op_142_cast_fp16")]; tensor var_143_split_sizes_0 = const()[name = string("op_143_split_sizes_0"), val = tensor([32, 32])]; int32 var_143_axis_0 = const()[name = string("op_143_axis_0"), val = int32(-2)]; tensor var_143_cast_fp16_0, tensor var_143_cast_fp16_1 = split(axis = var_143_axis_0, split_sizes = var_143_split_sizes_0, x = embed_1_cast_fp16)[name = string("op_143_cast_fp16")]; fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_145_cast_fp16 = mul(x = var_143_cast_fp16_1, y = const_2_promoted_to_fp16)[name = string("op_145_cast_fp16")]; bool var_147_interleave_0 = const()[name = string("op_147_interleave_0"), val = bool(false)]; tensor var_147_cast_fp16 = concat(axis = var_86, interleave = var_147_interleave_0, values = (var_145_cast_fp16, var_143_cast_fp16_0))[name = string("op_147_cast_fp16")]; tensor sin_to_fp16 = const()[name = string("sin_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2761664)))]; tensor var_148_cast_fp16 = mul(x = var_147_cast_fp16, y = sin_to_fp16)[name = string("op_148_cast_fp16")]; tensor query_states_3_cast_fp16 = add(x = var_142_cast_fp16, y = var_148_cast_fp16)[name = string("query_states_3_cast_fp16")]; tensor var_150_cast_fp16 = mul(x = embed_3_cast_fp16, y = cos_to_fp16)[name = string("op_150_cast_fp16")]; tensor var_151_split_sizes_0 = const()[name = string("op_151_split_sizes_0"), val = tensor([32, 32])]; int32 var_151_axis_0 = const()[name = string("op_151_axis_0"), val = int32(-2)]; tensor var_151_cast_fp16_0, tensor var_151_cast_fp16_1 = split(axis = var_151_axis_0, split_sizes = var_151_split_sizes_0, x = embed_3_cast_fp16)[name = string("op_151_cast_fp16")]; fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_153_cast_fp16 = mul(x = var_151_cast_fp16_1, y = const_3_promoted_to_fp16)[name = string("op_153_cast_fp16")]; bool var_155_interleave_0 = const()[name = string("op_155_interleave_0"), val = bool(false)]; tensor var_155_cast_fp16 = concat(axis = var_86, interleave = var_155_interleave_0, values = (var_153_cast_fp16, var_151_cast_fp16_0))[name = string("op_155_cast_fp16")]; tensor var_156_cast_fp16 = mul(x = var_155_cast_fp16, y = sin_to_fp16)[name = string("op_156_cast_fp16")]; tensor key_states_3_cast_fp16 = add(x = var_150_cast_fp16, y = var_156_cast_fp16)[name = string("key_states_3_cast_fp16")]; tensor var_161_split_sizes_0 = const()[name = string("op_161_split_sizes_0"), val = tensor([8, 8])]; int32 var_161_axis_0 = const()[name = string("op_161_axis_0"), val = int32(1)]; tensor var_161_cast_fp16_0, tensor var_161_cast_fp16_1 = split(axis = var_161_axis_0, split_sizes = var_161_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_161_cast_fp16")]; tensor var_163_split_sizes_0 = const()[name = string("op_163_split_sizes_0"), val = tensor([1, 1])]; int32 var_163_axis_0 = const()[name = string("op_163_axis_0"), val = int32(1)]; tensor var_163_cast_fp16_0, tensor var_163_cast_fp16_1 = split(axis = var_163_axis_0, split_sizes = var_163_split_sizes_0, x = key_states_3_cast_fp16)[name = string("op_163_cast_fp16")]; tensor var_165_split_sizes_0 = const()[name = string("op_165_split_sizes_0"), val = tensor([1, 1])]; int32 var_165_axis_0 = const()[name = string("op_165_axis_0"), val = int32(1)]; tensor var_165_cast_fp16_0, tensor var_165_cast_fp16_1 = split(axis = var_165_axis_0, split_sizes = var_165_split_sizes_0, x = value_states_3_cast_fp16)[name = string("op_165_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 attn_weights_1_cast_fp16 = matmul(transpose_x = attn_weights_1_transpose_x_1, transpose_y = attn_weights_1_transpose_y_1, x = var_163_cast_fp16_0, y = var_161_cast_fp16_0)[name = string("attn_weights_1_cast_fp16")]; fp16 _inversed_attn_weights_3_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_3_y_0_to_fp16"), val = fp16(0x1p-3)]; tensor _inversed_attn_weights_3_cast_fp16 = mul(x = attn_weights_1_cast_fp16, y = _inversed_attn_weights_3_y_0_to_fp16)[name = string("_inversed_attn_weights_3_cast_fp16")]; tensor attn_weights_5_cast_fp16 = softmax(axis = var_95, x = _inversed_attn_weights_3_cast_fp16)[name = string("attn_weights_5_cast_fp16")]; bool var_172_transpose_x_0 = const()[name = string("op_172_transpose_x_0"), val = bool(false)]; bool var_172_transpose_y_0 = const()[name = string("op_172_transpose_y_0"), val = bool(false)]; tensor var_172_cast_fp16 = matmul(transpose_x = var_172_transpose_x_0, transpose_y = var_172_transpose_y_0, x = var_165_cast_fp16_0, y = attn_weights_5_cast_fp16)[name = string("op_172_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 attn_weights_7_cast_fp16 = matmul(transpose_x = attn_weights_7_transpose_x_1, transpose_y = attn_weights_7_transpose_y_1, x = var_163_cast_fp16_1, y = var_161_cast_fp16_1)[name = string("attn_weights_7_cast_fp16")]; fp16 _inversed_attn_weights_9_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_9_y_0_to_fp16"), val = fp16(0x1p-3)]; tensor _inversed_attn_weights_9_cast_fp16 = mul(x = attn_weights_7_cast_fp16, y = _inversed_attn_weights_9_y_0_to_fp16)[name = string("_inversed_attn_weights_9_cast_fp16")]; tensor attn_weights_11_cast_fp16 = softmax(axis = var_95, x = _inversed_attn_weights_9_cast_fp16)[name = string("attn_weights_11_cast_fp16")]; bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)]; bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)]; tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = var_165_cast_fp16_1, y = attn_weights_11_cast_fp16)[name = string("attn_output_1_cast_fp16")]; bool attn_output_3_interleave_0 = const()[name = string("attn_output_3_interleave_0"), val = bool(false)]; tensor attn_output_3_cast_fp16 = concat(axis = var_90, interleave = attn_output_3_interleave_0, values = (var_172_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")]; tensor concat_3x = const()[name = string("concat_3x"), val = tensor([-1, 1024, 1, 3])]; tensor x_11_cast_fp16 = reshape(shape = concat_3x, x = attn_output_3_cast_fp16)[name = string("x_11_cast_fp16")]; string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")]; tensor hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor([1, 1])]; tensor hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor([0, 0, 0, 0])]; tensor hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor([1, 1])]; int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)]; tensor var_89_to_fp16 = const()[name = string("op_89_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2762112)))]; tensor hidden_states_3_cast_fp16 = conv(dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = var_89_to_fp16, x = x_11_cast_fp16)[name = string("hidden_states_3_cast_fp16")]; tensor x_13_cast_fp16 = add(x = x_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = string("x_13_cast_fp16")]; fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_191_cast_fp16 = mul(x = x_13_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_191_cast_fp16")]; bool x_15_interleave_0 = const()[name = string("x_15_interleave_0"), val = bool(false)]; tensor x_15_cast_fp16 = concat(axis = var_90, interleave = x_15_interleave_0, values = (x_13_cast_fp16, var_191_cast_fp16))[name = string("x_15_cast_fp16")]; tensor out_7_axes_0 = const()[name = string("out_7_axes_0"), val = tensor([1])]; fp16 var_201_to_fp16 = const()[name = string("op_201_to_fp16"), val = fp16(0x1.5p-17)]; tensor out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_201_to_fp16, x = x_15_cast_fp16)[name = string("out_7_cast_fp16")]; tensor layer_encoder_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4859328)))]; tensor out_9_cast_fp16 = mul(x = out_7_cast_fp16, y = layer_encoder_layers_0_post_attention_layernorm_weight_to_fp16)[name = string("out_9_cast_fp16")]; tensor var_207_split_sizes_0 = const()[name = string("op_207_split_sizes_0"), val = tensor([1024, 1024])]; int32 var_207_axis_0 = const()[name = string("op_207_axis_0"), val = int32(1)]; tensor var_207_cast_fp16_0, tensor var_207_cast_fp16_1 = split(axis = var_207_axis_0, split_sizes = var_207_split_sizes_0, x = out_9_cast_fp16)[name = string("op_207_cast_fp16")]; string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")]; tensor input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor([1, 1])]; tensor input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor([1, 1])]; int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)]; tensor var_76_to_fp16 = const()[name = string("op_76_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4863488)))]; tensor 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_76_to_fp16, x = var_207_cast_fp16_0)[name = string("input_1_cast_fp16")]; tensor var_215_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_215_cast_fp16")]; string var_220_pad_type_0 = const()[name = string("op_220_pad_type_0"), val = string("valid")]; tensor var_220_strides_0 = const()[name = string("op_220_strides_0"), val = tensor([1, 1])]; tensor var_220_pad_0 = const()[name = string("op_220_pad_0"), val = tensor([0, 0, 0, 0])]; tensor var_220_dilations_0 = const()[name = string("op_220_dilations_0"), val = tensor([1, 1])]; int32 var_220_groups_0 = const()[name = string("op_220_groups_0"), val = int32(1)]; tensor var_77_to_fp16 = const()[name = string("op_77_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13252160)))]; tensor var_220_cast_fp16 = conv(dilations = var_220_dilations_0, groups = var_220_groups_0, pad = var_220_pad_0, pad_type = var_220_pad_type_0, strides = var_220_strides_0, weight = var_77_to_fp16, x = var_207_cast_fp16_0)[name = string("op_220_cast_fp16")]; tensor x_21_cast_fp16 = mul(x = var_215_cast_fp16, y = var_220_cast_fp16)[name = string("x_21_cast_fp16")]; string hidden_states_5_pad_type_0 = const()[name = string("hidden_states_5_pad_type_0"), val = string("valid")]; tensor hidden_states_5_strides_0 = const()[name = string("hidden_states_5_strides_0"), val = tensor([1, 1])]; tensor hidden_states_5_pad_0 = const()[name = string("hidden_states_5_pad_0"), val = tensor([0, 0, 0, 0])]; tensor hidden_states_5_dilations_0 = const()[name = string("hidden_states_5_dilations_0"), val = tensor([1, 1])]; int32 hidden_states_5_groups_0 = const()[name = string("hidden_states_5_groups_0"), val = int32(1)]; tensor var_78_to_fp16 = const()[name = string("op_78_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21640832)))]; tensor 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_78_to_fp16, x = x_21_cast_fp16)[name = string("hidden_states_5_cast_fp16")]; tensor x_23_cast_fp16 = add(x = x_13_cast_fp16, y = hidden_states_5_cast_fp16)[name = string("x_23_cast_fp16")]; int32 var_238 = const()[name = string("op_238"), val = int32(-2)]; int32 var_242 = const()[name = string("op_242"), val = int32(1)]; int32 var_247 = const()[name = string("op_247"), val = int32(2)]; fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_252_cast_fp16 = mul(x = x_23_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_252_cast_fp16")]; bool x_25_interleave_0 = const()[name = string("x_25_interleave_0"), val = bool(false)]; tensor x_25_cast_fp16 = concat(axis = var_242, interleave = x_25_interleave_0, values = (x_23_cast_fp16, var_252_cast_fp16))[name = string("x_25_cast_fp16")]; tensor out_13_axes_0 = const()[name = string("out_13_axes_0"), val = tensor([1])]; fp16 var_262_to_fp16 = const()[name = string("op_262_to_fp16"), val = fp16(0x1.5p-17)]; tensor out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_262_to_fp16, x = x_25_cast_fp16)[name = string("out_13_cast_fp16")]; tensor layer_encoder_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_1_input_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30029504)))]; tensor out_15_cast_fp16 = mul(x = out_13_cast_fp16, y = layer_encoder_layers_1_input_layernorm_weight_to_fp16)[name = string("out_15_cast_fp16")]; tensor var_268_split_sizes_0 = const()[name = string("op_268_split_sizes_0"), val = tensor([1024, 1024])]; int32 var_268_axis_0 = const()[name = string("op_268_axis_0"), val = int32(1)]; tensor var_268_cast_fp16_0, tensor var_268_cast_fp16_1 = split(axis = var_268_axis_0, split_sizes = var_268_split_sizes_0, x = out_15_cast_fp16)[name = string("op_268_cast_fp16")]; string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")]; tensor query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor([1, 1])]; tensor query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor([0, 0, 0, 0])]; tensor query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor([1, 1])]; int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)]; tensor var_233_to_fp16 = const()[name = string("op_233_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30033664)))]; tensor query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = var_233_to_fp16, x = var_268_cast_fp16_0)[name = string("query_states_7_cast_fp16")]; string key_states_7_pad_type_0 = const()[name = string("key_states_7_pad_type_0"), val = string("valid")]; tensor key_states_7_strides_0 = const()[name = string("key_states_7_strides_0"), val = tensor([1, 1])]; tensor key_states_7_pad_0 = const()[name = string("key_states_7_pad_0"), val = tensor([0, 0, 0, 0])]; tensor key_states_7_dilations_0 = const()[name = string("key_states_7_dilations_0"), val = tensor([1, 1])]; int32 key_states_7_groups_0 = const()[name = string("key_states_7_groups_0"), val = int32(1)]; tensor var_234_to_fp16 = const()[name = string("op_234_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32130880)))]; tensor key_states_7_cast_fp16 = conv(dilations = key_states_7_dilations_0, groups = key_states_7_groups_0, pad = key_states_7_pad_0, pad_type = key_states_7_pad_type_0, strides = key_states_7_strides_0, weight = var_234_to_fp16, x = var_268_cast_fp16_0)[name = string("key_states_7_cast_fp16")]; string value_states_7_pad_type_0 = const()[name = string("value_states_7_pad_type_0"), val = string("valid")]; tensor value_states_7_strides_0 = const()[name = string("value_states_7_strides_0"), val = tensor([1, 1])]; tensor value_states_7_pad_0 = const()[name = string("value_states_7_pad_0"), val = tensor([0, 0, 0, 0])]; tensor value_states_7_dilations_0 = const()[name = string("value_states_7_dilations_0"), val = tensor([1, 1])]; int32 value_states_7_groups_0 = const()[name = string("value_states_7_groups_0"), val = int32(1)]; tensor var_235_to_fp16 = const()[name = string("op_235_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32393088)))]; tensor value_states_7_cast_fp16 = conv(dilations = value_states_7_dilations_0, groups = value_states_7_groups_0, pad = value_states_7_pad_0, pad_type = value_states_7_pad_type_0, strides = value_states_7_strides_0, weight = var_235_to_fp16, x = var_268_cast_fp16_0)[name = string("value_states_7_cast_fp16")]; tensor concat_4x = const()[name = string("concat_4x"), val = tensor([-1, 16, 64, 3])]; tensor embed_5_cast_fp16 = reshape(shape = concat_4x, x = query_states_7_cast_fp16)[name = string("embed_5_cast_fp16")]; tensor concat_5x = const()[name = string("concat_5x"), val = tensor([-1, 2, 64, 3])]; tensor embed_7_cast_fp16 = reshape(shape = concat_5x, x = key_states_7_cast_fp16)[name = string("embed_7_cast_fp16")]; tensor concat_6x = const()[name = string("concat_6x"), val = tensor([-1, 2, 64, 3])]; tensor value_states_9_cast_fp16 = reshape(shape = concat_6x, x = value_states_7_cast_fp16)[name = string("value_states_9_cast_fp16")]; tensor var_294_cast_fp16 = mul(x = embed_5_cast_fp16, y = cos_to_fp16)[name = string("op_294_cast_fp16")]; tensor var_295_split_sizes_0 = const()[name = string("op_295_split_sizes_0"), val = tensor([32, 32])]; int32 var_295_axis_0 = const()[name = string("op_295_axis_0"), val = int32(-2)]; tensor var_295_cast_fp16_0, tensor var_295_cast_fp16_1 = split(axis = var_295_axis_0, split_sizes = var_295_split_sizes_0, x = embed_5_cast_fp16)[name = string("op_295_cast_fp16")]; fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_297_cast_fp16 = mul(x = var_295_cast_fp16_1, y = const_6_promoted_to_fp16)[name = string("op_297_cast_fp16")]; bool var_299_interleave_0 = const()[name = string("op_299_interleave_0"), val = bool(false)]; tensor var_299_cast_fp16 = concat(axis = var_238, interleave = var_299_interleave_0, values = (var_297_cast_fp16, var_295_cast_fp16_0))[name = string("op_299_cast_fp16")]; tensor var_300_cast_fp16 = mul(x = var_299_cast_fp16, y = sin_to_fp16)[name = string("op_300_cast_fp16")]; tensor query_states_9_cast_fp16 = add(x = var_294_cast_fp16, y = var_300_cast_fp16)[name = string("query_states_9_cast_fp16")]; tensor var_302_cast_fp16 = mul(x = embed_7_cast_fp16, y = cos_to_fp16)[name = string("op_302_cast_fp16")]; tensor var_303_split_sizes_0 = const()[name = string("op_303_split_sizes_0"), val = tensor([32, 32])]; int32 var_303_axis_0 = const()[name = string("op_303_axis_0"), val = int32(-2)]; tensor var_303_cast_fp16_0, tensor var_303_cast_fp16_1 = split(axis = var_303_axis_0, split_sizes = var_303_split_sizes_0, x = embed_7_cast_fp16)[name = string("op_303_cast_fp16")]; fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_305_cast_fp16 = mul(x = var_303_cast_fp16_1, y = const_7_promoted_to_fp16)[name = string("op_305_cast_fp16")]; bool var_307_interleave_0 = const()[name = string("op_307_interleave_0"), val = bool(false)]; tensor var_307_cast_fp16 = concat(axis = var_238, interleave = var_307_interleave_0, values = (var_305_cast_fp16, var_303_cast_fp16_0))[name = string("op_307_cast_fp16")]; tensor var_308_cast_fp16 = mul(x = var_307_cast_fp16, y = sin_to_fp16)[name = string("op_308_cast_fp16")]; tensor key_states_9_cast_fp16 = add(x = var_302_cast_fp16, y = var_308_cast_fp16)[name = string("key_states_9_cast_fp16")]; tensor var_313_split_sizes_0 = const()[name = string("op_313_split_sizes_0"), val = tensor([8, 8])]; int32 var_313_axis_0 = const()[name = string("op_313_axis_0"), val = int32(1)]; tensor var_313_cast_fp16_0, tensor var_313_cast_fp16_1 = split(axis = var_313_axis_0, split_sizes = var_313_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_313_cast_fp16")]; tensor var_315_split_sizes_0 = const()[name = string("op_315_split_sizes_0"), val = tensor([1, 1])]; int32 var_315_axis_0 = const()[name = string("op_315_axis_0"), val = int32(1)]; tensor var_315_cast_fp16_0, tensor var_315_cast_fp16_1 = split(axis = var_315_axis_0, split_sizes = var_315_split_sizes_0, x = key_states_9_cast_fp16)[name = string("op_315_cast_fp16")]; tensor var_317_split_sizes_0 = const()[name = string("op_317_split_sizes_0"), val = tensor([1, 1])]; int32 var_317_axis_0 = const()[name = string("op_317_axis_0"), val = int32(1)]; tensor var_317_cast_fp16_0, tensor var_317_cast_fp16_1 = split(axis = var_317_axis_0, split_sizes = var_317_split_sizes_0, x = value_states_9_cast_fp16)[name = string("op_317_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 attn_weights_13_cast_fp16 = matmul(transpose_x = attn_weights_13_transpose_x_1, transpose_y = attn_weights_13_transpose_y_1, x = var_315_cast_fp16_0, y = var_313_cast_fp16_0)[name = string("attn_weights_13_cast_fp16")]; fp16 _inversed_attn_weights_15_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_15_y_0_to_fp16"), val = fp16(0x1p-3)]; tensor _inversed_attn_weights_15_cast_fp16 = mul(x = attn_weights_13_cast_fp16, y = _inversed_attn_weights_15_y_0_to_fp16)[name = string("_inversed_attn_weights_15_cast_fp16")]; tensor attn_weights_17_cast_fp16 = softmax(axis = var_247, x = _inversed_attn_weights_15_cast_fp16)[name = string("attn_weights_17_cast_fp16")]; bool var_324_transpose_x_0 = const()[name = string("op_324_transpose_x_0"), val = bool(false)]; bool var_324_transpose_y_0 = const()[name = string("op_324_transpose_y_0"), val = bool(false)]; tensor var_324_cast_fp16 = matmul(transpose_x = var_324_transpose_x_0, transpose_y = var_324_transpose_y_0, x = var_317_cast_fp16_0, y = attn_weights_17_cast_fp16)[name = string("op_324_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 attn_weights_19_cast_fp16 = matmul(transpose_x = attn_weights_19_transpose_x_1, transpose_y = attn_weights_19_transpose_y_1, x = var_315_cast_fp16_1, y = var_313_cast_fp16_1)[name = string("attn_weights_19_cast_fp16")]; fp16 _inversed_attn_weights_21_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_21_y_0_to_fp16"), val = fp16(0x1p-3)]; tensor _inversed_attn_weights_21_cast_fp16 = mul(x = attn_weights_19_cast_fp16, y = _inversed_attn_weights_21_y_0_to_fp16)[name = string("_inversed_attn_weights_21_cast_fp16")]; tensor attn_weights_23_cast_fp16 = softmax(axis = var_247, x = _inversed_attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")]; bool attn_output_5_transpose_x_0 = const()[name = string("attn_output_5_transpose_x_0"), val = bool(false)]; bool attn_output_5_transpose_y_0 = const()[name = string("attn_output_5_transpose_y_0"), val = bool(false)]; tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = var_317_cast_fp16_1, y = attn_weights_23_cast_fp16)[name = string("attn_output_5_cast_fp16")]; bool attn_output_7_interleave_0 = const()[name = string("attn_output_7_interleave_0"), val = bool(false)]; tensor attn_output_7_cast_fp16 = concat(axis = var_242, interleave = attn_output_7_interleave_0, values = (var_324_cast_fp16, attn_output_5_cast_fp16))[name = string("attn_output_7_cast_fp16")]; tensor concat_7x = const()[name = string("concat_7x"), val = tensor([-1, 1024, 1, 3])]; tensor x_29_cast_fp16 = reshape(shape = concat_7x, x = attn_output_7_cast_fp16)[name = string("x_29_cast_fp16")]; string hidden_states_9_pad_type_0 = const()[name = string("hidden_states_9_pad_type_0"), val = string("valid")]; tensor hidden_states_9_strides_0 = const()[name = string("hidden_states_9_strides_0"), val = tensor([1, 1])]; tensor hidden_states_9_pad_0 = const()[name = string("hidden_states_9_pad_0"), val = tensor([0, 0, 0, 0])]; tensor hidden_states_9_dilations_0 = const()[name = string("hidden_states_9_dilations_0"), val = tensor([1, 1])]; int32 hidden_states_9_groups_0 = const()[name = string("hidden_states_9_groups_0"), val = int32(1)]; tensor var_241_to_fp16 = const()[name = string("op_241_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32655296)))]; tensor 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_241_to_fp16, x = x_29_cast_fp16)[name = string("hidden_states_9_cast_fp16")]; tensor x_31_cast_fp16 = add(x = x_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = string("x_31_cast_fp16")]; fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_343_cast_fp16 = mul(x = x_31_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_343_cast_fp16")]; bool x_33_interleave_0 = const()[name = string("x_33_interleave_0"), val = bool(false)]; tensor x_33_cast_fp16 = concat(axis = var_242, interleave = x_33_interleave_0, values = (x_31_cast_fp16, var_343_cast_fp16))[name = string("x_33_cast_fp16")]; tensor out_19_axes_0 = const()[name = string("out_19_axes_0"), val = tensor([1])]; fp16 var_353_to_fp16 = const()[name = string("op_353_to_fp16"), val = fp16(0x1.5p-17)]; tensor out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_353_to_fp16, x = x_33_cast_fp16)[name = string("out_19_cast_fp16")]; tensor layer_encoder_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34752512)))]; tensor out_21_cast_fp16 = mul(x = out_19_cast_fp16, y = layer_encoder_layers_1_post_attention_layernorm_weight_to_fp16)[name = string("out_21_cast_fp16")]; tensor var_359_split_sizes_0 = const()[name = string("op_359_split_sizes_0"), val = tensor([1024, 1024])]; int32 var_359_axis_0 = const()[name = string("op_359_axis_0"), val = int32(1)]; tensor var_359_cast_fp16_0, tensor var_359_cast_fp16_1 = split(axis = var_359_axis_0, split_sizes = var_359_split_sizes_0, x = out_21_cast_fp16)[name = string("op_359_cast_fp16")]; string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")]; tensor input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor([1, 1])]; tensor input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor([1, 1])]; int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)]; tensor var_228_to_fp16 = const()[name = string("op_228_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34756672)))]; tensor 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_228_to_fp16, x = var_359_cast_fp16_0)[name = string("input_3_cast_fp16")]; tensor var_367_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_367_cast_fp16")]; string var_372_pad_type_0 = const()[name = string("op_372_pad_type_0"), val = string("valid")]; tensor var_372_strides_0 = const()[name = string("op_372_strides_0"), val = tensor([1, 1])]; tensor var_372_pad_0 = const()[name = string("op_372_pad_0"), val = tensor([0, 0, 0, 0])]; tensor var_372_dilations_0 = const()[name = string("op_372_dilations_0"), val = tensor([1, 1])]; int32 var_372_groups_0 = const()[name = string("op_372_groups_0"), val = int32(1)]; tensor var_229_to_fp16 = const()[name = string("op_229_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43145344)))]; tensor var_372_cast_fp16 = conv(dilations = var_372_dilations_0, groups = var_372_groups_0, pad = var_372_pad_0, pad_type = var_372_pad_type_0, strides = var_372_strides_0, weight = var_229_to_fp16, x = var_359_cast_fp16_0)[name = string("op_372_cast_fp16")]; tensor x_39_cast_fp16 = mul(x = var_367_cast_fp16, y = var_372_cast_fp16)[name = string("x_39_cast_fp16")]; string hidden_states_11_pad_type_0 = const()[name = string("hidden_states_11_pad_type_0"), val = string("valid")]; tensor hidden_states_11_strides_0 = const()[name = string("hidden_states_11_strides_0"), val = tensor([1, 1])]; tensor hidden_states_11_pad_0 = const()[name = string("hidden_states_11_pad_0"), val = tensor([0, 0, 0, 0])]; tensor hidden_states_11_dilations_0 = const()[name = string("hidden_states_11_dilations_0"), val = tensor([1, 1])]; int32 hidden_states_11_groups_0 = const()[name = string("hidden_states_11_groups_0"), val = int32(1)]; tensor var_230_to_fp16 = const()[name = string("op_230_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51534016)))]; tensor hidden_states_11_cast_fp16 = conv(dilations = hidden_states_11_dilations_0, groups = hidden_states_11_groups_0, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = hidden_states_11_strides_0, weight = var_230_to_fp16, x = x_39_cast_fp16)[name = string("hidden_states_11_cast_fp16")]; tensor x_41_cast_fp16 = add(x = x_31_cast_fp16, y = hidden_states_11_cast_fp16)[name = string("x_41_cast_fp16")]; int32 var_390 = const()[name = string("op_390"), val = int32(-2)]; int32 var_394 = const()[name = string("op_394"), val = int32(1)]; int32 var_399 = const()[name = string("op_399"), val = int32(2)]; fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_404_cast_fp16 = mul(x = x_41_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_404_cast_fp16")]; bool x_43_interleave_0 = const()[name = string("x_43_interleave_0"), val = bool(false)]; tensor x_43_cast_fp16 = concat(axis = var_394, interleave = x_43_interleave_0, values = (x_41_cast_fp16, var_404_cast_fp16))[name = string("x_43_cast_fp16")]; tensor out_25_axes_0 = const()[name = string("out_25_axes_0"), val = tensor([1])]; fp16 var_414_to_fp16 = const()[name = string("op_414_to_fp16"), val = fp16(0x1.5p-17)]; tensor out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_414_to_fp16, x = x_43_cast_fp16)[name = string("out_25_cast_fp16")]; tensor layer_encoder_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("layer_encoder_layers_2_input_layernorm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59922688)))]; tensor out_27_cast_fp16 = mul(x = out_25_cast_fp16, y = layer_encoder_layers_2_input_layernorm_weight_to_fp16)[name = string("out_27_cast_fp16")]; tensor var_420_split_sizes_0 = const()[name = string("op_420_split_sizes_0"), val = tensor([1024, 1024])]; int32 var_420_axis_0 = const()[name = string("op_420_axis_0"), val = int32(1)]; tensor var_420_cast_fp16_0, tensor var_420_cast_fp16_1 = split(axis = var_420_axis_0, split_sizes = var_420_split_sizes_0, x = out_27_cast_fp16)[name = string("op_420_cast_fp16")]; string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")]; tensor query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor([1, 1])]; tensor query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor([0, 0, 0, 0])]; tensor query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor([1, 1])]; int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)]; tensor var_385_to_fp16 = const()[name = string("op_385_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59926848)))]; tensor 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_385_to_fp16, x = var_420_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 key_states_13_strides_0 = const()[name = string("key_states_13_strides_0"), val = tensor([1, 1])]; tensor key_states_13_pad_0 = const()[name = string("key_states_13_pad_0"), val = tensor([0, 0, 0, 0])]; tensor key_states_13_dilations_0 = const()[name = string("key_states_13_dilations_0"), val = tensor([1, 1])]; int32 key_states_13_groups_0 = const()[name = string("key_states_13_groups_0"), val = int32(1)]; tensor var_386_to_fp16 = const()[name = string("op_386_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62024064)))]; tensor 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_386_to_fp16, x = var_420_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 value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor([1, 1])]; tensor value_states_13_pad_0 = const()[name = string("value_states_13_pad_0"), val = tensor([0, 0, 0, 0])]; tensor value_states_13_dilations_0 = const()[name = string("value_states_13_dilations_0"), val = tensor([1, 1])]; int32 value_states_13_groups_0 = const()[name = string("value_states_13_groups_0"), val = int32(1)]; tensor var_387_to_fp16 = const()[name = string("op_387_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62286272)))]; tensor 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_387_to_fp16, x = var_420_cast_fp16_0)[name = string("value_states_13_cast_fp16")]; tensor concat_8x = const()[name = string("concat_8x"), val = tensor([-1, 16, 64, 3])]; tensor embed_9_cast_fp16 = reshape(shape = concat_8x, x = query_states_13_cast_fp16)[name = string("embed_9_cast_fp16")]; tensor concat_9x = const()[name = string("concat_9x"), val = tensor([-1, 2, 64, 3])]; tensor embed_11_cast_fp16 = reshape(shape = concat_9x, x = key_states_13_cast_fp16)[name = string("embed_11_cast_fp16")]; tensor concat_10x = const()[name = string("concat_10x"), val = tensor([-1, 2, 64, 3])]; tensor value_states_15_cast_fp16 = reshape(shape = concat_10x, x = value_states_13_cast_fp16)[name = string("value_states_15_cast_fp16")]; tensor var_446_cast_fp16 = mul(x = embed_9_cast_fp16, y = cos_to_fp16)[name = string("op_446_cast_fp16")]; tensor var_447_split_sizes_0 = const()[name = string("op_447_split_sizes_0"), val = tensor([32, 32])]; int32 var_447_axis_0 = const()[name = string("op_447_axis_0"), val = int32(-2)]; tensor var_447_cast_fp16_0, tensor var_447_cast_fp16_1 = split(axis = var_447_axis_0, split_sizes = var_447_split_sizes_0, x = embed_9_cast_fp16)[name = string("op_447_cast_fp16")]; fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_449_cast_fp16 = mul(x = var_447_cast_fp16_1, y = const_10_promoted_to_fp16)[name = string("op_449_cast_fp16")]; bool var_451_interleave_0 = const()[name = string("op_451_interleave_0"), val = bool(false)]; tensor var_451_cast_fp16 = concat(axis = var_390, interleave = var_451_interleave_0, values = (var_449_cast_fp16, var_447_cast_fp16_0))[name = string("op_451_cast_fp16")]; tensor var_452_cast_fp16 = mul(x = var_451_cast_fp16, y = sin_to_fp16)[name = string("op_452_cast_fp16")]; tensor query_states_15_cast_fp16 = add(x = var_446_cast_fp16, y = var_452_cast_fp16)[name = string("query_states_15_cast_fp16")]; tensor var_454_cast_fp16 = mul(x = embed_11_cast_fp16, y = cos_to_fp16)[name = string("op_454_cast_fp16")]; tensor var_455_split_sizes_0 = const()[name = string("op_455_split_sizes_0"), val = tensor([32, 32])]; int32 var_455_axis_0 = const()[name = string("op_455_axis_0"), val = int32(-2)]; tensor var_455_cast_fp16_0, tensor var_455_cast_fp16_1 = split(axis = var_455_axis_0, split_sizes = var_455_split_sizes_0, x = embed_11_cast_fp16)[name = string("op_455_cast_fp16")]; fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_457_cast_fp16 = mul(x = var_455_cast_fp16_1, y = const_11_promoted_to_fp16)[name = string("op_457_cast_fp16")]; bool var_459_interleave_0 = const()[name = string("op_459_interleave_0"), val = bool(false)]; tensor var_459_cast_fp16 = concat(axis = var_390, interleave = var_459_interleave_0, values = (var_457_cast_fp16, var_455_cast_fp16_0))[name = string("op_459_cast_fp16")]; tensor var_460_cast_fp16 = mul(x = var_459_cast_fp16, y = sin_to_fp16)[name = string("op_460_cast_fp16")]; tensor key_states_15_cast_fp16 = add(x = var_454_cast_fp16, y = var_460_cast_fp16)[name = string("key_states_15_cast_fp16")]; tensor var_465_split_sizes_0 = const()[name = string("op_465_split_sizes_0"), val = tensor([8, 8])]; int32 var_465_axis_0 = const()[name = string("op_465_axis_0"), val = int32(1)]; tensor var_465_cast_fp16_0, tensor var_465_cast_fp16_1 = split(axis = var_465_axis_0, split_sizes = var_465_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_465_cast_fp16")]; tensor var_467_split_sizes_0 = const()[name = string("op_467_split_sizes_0"), val = tensor([1, 1])]; int32 var_467_axis_0 = const()[name = string("op_467_axis_0"), val = int32(1)]; tensor var_467_cast_fp16_0, tensor var_467_cast_fp16_1 = split(axis = var_467_axis_0, split_sizes = var_467_split_sizes_0, x = key_states_15_cast_fp16)[name = string("op_467_cast_fp16")]; tensor var_469_split_sizes_0 = const()[name = string("op_469_split_sizes_0"), val = tensor([1, 1])]; int32 var_469_axis_0 = const()[name = string("op_469_axis_0"), val = int32(1)]; tensor var_469_cast_fp16_0, tensor var_469_cast_fp16_1 = split(axis = var_469_axis_0, split_sizes = var_469_split_sizes_0, x = value_states_15_cast_fp16)[name = string("op_469_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 attn_weights_25_cast_fp16 = matmul(transpose_x = attn_weights_25_transpose_x_1, transpose_y = attn_weights_25_transpose_y_1, x = var_467_cast_fp16_0, y = var_465_cast_fp16_0)[name = string("attn_weights_25_cast_fp16")]; fp16 _inversed_attn_weights_27_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_27_y_0_to_fp16"), val = fp16(0x1p-3)]; tensor _inversed_attn_weights_27_cast_fp16 = mul(x = attn_weights_25_cast_fp16, y = _inversed_attn_weights_27_y_0_to_fp16)[name = string("_inversed_attn_weights_27_cast_fp16")]; tensor attn_weights_29_cast_fp16 = softmax(axis = var_399, x = _inversed_attn_weights_27_cast_fp16)[name = string("attn_weights_29_cast_fp16")]; bool var_476_transpose_x_0 = const()[name = string("op_476_transpose_x_0"), val = bool(false)]; bool var_476_transpose_y_0 = const()[name = string("op_476_transpose_y_0"), val = bool(false)]; tensor var_476_cast_fp16 = matmul(transpose_x = var_476_transpose_x_0, transpose_y = var_476_transpose_y_0, x = var_469_cast_fp16_0, y = attn_weights_29_cast_fp16)[name = string("op_476_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 attn_weights_31_cast_fp16 = matmul(transpose_x = attn_weights_31_transpose_x_1, transpose_y = attn_weights_31_transpose_y_1, x = var_467_cast_fp16_1, y = var_465_cast_fp16_1)[name = string("attn_weights_31_cast_fp16")]; fp16 _inversed_attn_weights_33_y_0_to_fp16 = const()[name = string("_inversed_attn_weights_33_y_0_to_fp16"), val = fp16(0x1p-3)]; tensor _inversed_attn_weights_33_cast_fp16 = mul(x = attn_weights_31_cast_fp16, y = _inversed_attn_weights_33_y_0_to_fp16)[name = string("_inversed_attn_weights_33_cast_fp16")]; tensor