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, 4]}}), ("EnumeratedShapes", {{"2e922f8e", {{"x", [24, 64, 1, 4]}}}, {"45c08af7", {{"x", [30, 64, 1, 4]}}}, {"66582a1a", {{"x", [10, 64, 1, 4]}}}, {"77b3f94e", {{"x", [20, 64, 1, 4]}}}, {"7f4b84c7", {{"x", [26, 64, 1, 4]}}}, {"7fe2a293", {{"x", [16, 64, 1, 4]}}}, {"8523ce80", {{"x", [22, 64, 1, 4]}}}, {"92b5cb49", {{"x", [8, 64, 1, 4]}}}, {"a76a4363", {{"x", [32, 64, 1, 4]}}}, {"b23494e7", {{"x", [14, 64, 1, 4]}}}, {"ba7abd9b", {{"x", [18, 64, 1, 4]}}}, {"cb00930d", {{"x", [28, 64, 1, 4]}}}, {"df778eb5", {{"x", [12, 64, 1, 4]}}}, {"e3e45208", {{"x", [1, 64, 1, 4]}}}})))] { 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_75_to_fp16 = const()[name = string("op_75_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_75_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_92_begin_0 = const()[name = string("op_92_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_92_end_0 = const()[name = string("op_92_end_0"), val = tensor([0, 1024, 1, 1])]; tensor var_92_end_mask_0 = const()[name = string("op_92_end_mask_0"), val = tensor([true, true, true, false])]; tensor var_92_cast_fp16 = slice_by_index(begin = var_92_begin_0, end = var_92_end_0, end_mask = var_92_end_mask_0, x = fill_like_0_cast_fp16)[name = string("op_92_cast_fp16")]; tensor var_93_to_fp16 = const()[name = string("op_93_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133312)))]; tensor special_tokens_cast_fp16 = mul(x = var_92_cast_fp16, y = var_93_to_fp16)[name = string("special_tokens_cast_fp16")]; int32 var_96 = const()[name = string("op_96"), 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_96, interleave = x_5_interleave_0, values = (special_tokens_cast_fp16, x_3_cast_fp16))[name = string("x_5_cast_fp16")]; int32 var_108 = const()[name = string("op_108"), val = int32(-2)]; int32 var_112 = const()[name = string("op_112"), val = int32(1)]; int32 var_117 = const()[name = string("op_117"), val = int32(2)]; fp16 const_1_promoted_to_fp16 = const()[name = string("const_1_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_122_cast_fp16 = mul(x = x_5_cast_fp16, y = const_1_promoted_to_fp16)[name = string("op_122_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_112, interleave = x_7_interleave_0, values = (x_5_cast_fp16, var_122_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_132_to_fp16 = const()[name = string("op_132_to_fp16"), val = fp16(0x1.5p-17)]; tensor out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_132_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_138_split_sizes_0 = const()[name = string("op_138_split_sizes_0"), val = tensor([1024, 1024])]; int32 var_138_axis_0 = const()[name = string("op_138_axis_0"), val = int32(1)]; tensor var_138_cast_fp16_0, tensor var_138_cast_fp16_1 = split(axis = var_138_axis_0, split_sizes = var_138_split_sizes_0, x = out_3_cast_fp16)[name = string("op_138_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_103_to_fp16 = const()[name = string("op_103_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_103_to_fp16, x = var_138_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_104_to_fp16 = const()[name = string("op_104_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_104_to_fp16, x = var_138_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_105_to_fp16 = const()[name = string("op_105_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_105_to_fp16, x = var_138_cast_fp16_0)[name = string("value_states_1_cast_fp16")]; tensor concat_0x = const()[name = string("concat_0x"), val = tensor([-1, 16, 64, 5])]; 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, 5])]; 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, 5])]; 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_164_cast_fp16 = mul(x = embed_1_cast_fp16, y = cos_to_fp16)[name = string("op_164_cast_fp16")]; tensor var_165_split_sizes_0 = const()[name = string("op_165_split_sizes_0"), val = tensor([32, 32])]; int32 var_165_axis_0 = const()[name = string("op_165_axis_0"), val = int32(-2)]; 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 = embed_1_cast_fp16)[name = string("op_165_cast_fp16")]; fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_167_cast_fp16 = mul(x = var_165_cast_fp16_1, y = const_2_promoted_to_fp16)[name = string("op_167_cast_fp16")]; bool var_169_interleave_0 = const()[name = string("op_169_interleave_0"), val = bool(false)]; tensor var_169_cast_fp16 = concat(axis = var_108, interleave = var_169_interleave_0, values = (var_167_cast_fp16, var_165_cast_fp16_0))[name = string("op_169_cast_fp16")]; tensor sin_to_fp16 = const()[name = string("sin_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2761920)))]; tensor var_170_cast_fp16 = mul(x = var_169_cast_fp16, y = sin_to_fp16)[name = string("op_170_cast_fp16")]; tensor query_states_3_cast_fp16 = add(x = var_164_cast_fp16, y = var_170_cast_fp16)[name = string("query_states_3_cast_fp16")]; tensor var_172_cast_fp16 = mul(x = embed_3_cast_fp16, y = cos_to_fp16)[name = string("op_172_cast_fp16")]; tensor var_173_split_sizes_0 = const()[name = string("op_173_split_sizes_0"), val = tensor([32, 32])]; int32 var_173_axis_0 = const()[name = string("op_173_axis_0"), val = int32(-2)]; tensor var_173_cast_fp16_0, tensor var_173_cast_fp16_1 = split(axis = var_173_axis_0, split_sizes = var_173_split_sizes_0, x = embed_3_cast_fp16)[name = string("op_173_cast_fp16")]; fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_175_cast_fp16 = mul(x = var_173_cast_fp16_1, y = const_3_promoted_to_fp16)[name = string("op_175_cast_fp16")]; bool var_177_interleave_0 = const()[name = string("op_177_interleave_0"), val = bool(false)]; tensor var_177_cast_fp16 = concat(axis = var_108, interleave = var_177_interleave_0, values = (var_175_cast_fp16, var_173_cast_fp16_0))[name = string("op_177_cast_fp16")]; tensor var_178_cast_fp16 = mul(x = var_177_cast_fp16, y = sin_to_fp16)[name = string("op_178_cast_fp16")]; tensor key_states_3_cast_fp16 = add(x = var_172_cast_fp16, y = var_178_cast_fp16)[name = string("key_states_3_cast_fp16")]; tensor var_183_split_sizes_0 = const()[name = string("op_183_split_sizes_0"), val = tensor([8, 8])]; int32 var_183_axis_0 = const()[name = string("op_183_axis_0"), val = int32(1)]; tensor var_183_cast_fp16_0, tensor var_183_cast_fp16_1 = split(axis = var_183_axis_0, split_sizes = var_183_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_183_cast_fp16")]; tensor var_185_split_sizes_0 = const()[name = string("op_185_split_sizes_0"), val = tensor([1, 1])]; int32 var_185_axis_0 = const()[name = string("op_185_axis_0"), val = int32(1)]; tensor var_185_cast_fp16_0, tensor var_185_cast_fp16_1 = split(axis = var_185_axis_0, split_sizes = var_185_split_sizes_0, x = key_states_3_cast_fp16)[name = string("op_185_cast_fp16")]; tensor var_187_split_sizes_0 = const()[name = string("op_187_split_sizes_0"), val = tensor([1, 1])]; int32 var_187_axis_0 = const()[name = string("op_187_axis_0"), val = int32(1)]; tensor var_187_cast_fp16_0, tensor var_187_cast_fp16_1 = split(axis = var_187_axis_0, split_sizes = var_187_split_sizes_0, x = value_states_3_cast_fp16)[name = string("op_187_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_185_cast_fp16_0, y = var_183_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_117, x = _inversed_attn_weights_3_cast_fp16)[name = string("attn_weights_5_cast_fp16")]; bool var_194_transpose_x_0 = const()[name = string("op_194_transpose_x_0"), val = bool(false)]; bool var_194_transpose_y_0 = const()[name = string("op_194_transpose_y_0"), val = bool(false)]; tensor var_194_cast_fp16 = matmul(transpose_x = var_194_transpose_x_0, transpose_y = var_194_transpose_y_0, x = var_187_cast_fp16_0, y = attn_weights_5_cast_fp16)[name = string("op_194_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_185_cast_fp16_1, y = var_183_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_117, 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_187_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_112, interleave = attn_output_3_interleave_0, values = (var_194_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, 5])]; 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_111_to_fp16 = const()[name = string("op_111_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2762624)))]; 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_111_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_213_cast_fp16 = mul(x = x_13_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_213_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_112, interleave = x_15_interleave_0, values = (x_13_cast_fp16, var_213_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_223_to_fp16 = const()[name = string("op_223_to_fp16"), val = fp16(0x1.5p-17)]; tensor out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_223_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(4859840)))]; 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_229_split_sizes_0 = const()[name = string("op_229_split_sizes_0"), val = tensor([1024, 1024])]; int32 var_229_axis_0 = const()[name = string("op_229_axis_0"), val = int32(1)]; tensor var_229_cast_fp16_0, tensor var_229_cast_fp16_1 = split(axis = var_229_axis_0, split_sizes = var_229_split_sizes_0, x = out_9_cast_fp16)[name = string("op_229_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_98_to_fp16 = const()[name = string("op_98_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4864000)))]; 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_98_to_fp16, x = var_229_cast_fp16_0)[name = string("input_1_cast_fp16")]; tensor var_237_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_237_cast_fp16")]; string var_242_pad_type_0 = const()[name = string("op_242_pad_type_0"), val = string("valid")]; tensor var_242_strides_0 = const()[name = string("op_242_strides_0"), val = tensor([1, 1])]; tensor var_242_pad_0 = const()[name = string("op_242_pad_0"), val = tensor([0, 0, 0, 0])]; tensor var_242_dilations_0 = const()[name = string("op_242_dilations_0"), val = tensor([1, 1])]; int32 var_242_groups_0 = const()[name = string("op_242_groups_0"), val = int32(1)]; tensor var_99_to_fp16 = const()[name = string("op_99_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13252672)))]; tensor var_242_cast_fp16 = conv(dilations = var_242_dilations_0, groups = var_242_groups_0, pad = var_242_pad_0, pad_type = var_242_pad_type_0, strides = var_242_strides_0, weight = var_99_to_fp16, x = var_229_cast_fp16_0)[name = string("op_242_cast_fp16")]; tensor x_21_cast_fp16 = mul(x = var_237_cast_fp16, y = var_242_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_100_to_fp16 = const()[name = string("op_100_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21641344)))]; 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_100_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_260 = const()[name = string("op_260"), val = int32(-2)]; int32 var_264 = const()[name = string("op_264"), val = int32(1)]; int32 var_269 = const()[name = string("op_269"), val = int32(2)]; fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_274_cast_fp16 = mul(x = x_23_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_274_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_264, interleave = x_25_interleave_0, values = (x_23_cast_fp16, var_274_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_284_to_fp16 = const()[name = string("op_284_to_fp16"), val = fp16(0x1.5p-17)]; tensor out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_284_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(30030016)))]; 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_290_split_sizes_0 = const()[name = string("op_290_split_sizes_0"), val = tensor([1024, 1024])]; int32 var_290_axis_0 = const()[name = string("op_290_axis_0"), val = int32(1)]; tensor var_290_cast_fp16_0, tensor var_290_cast_fp16_1 = split(axis = var_290_axis_0, split_sizes = var_290_split_sizes_0, x = out_15_cast_fp16)[name = string("op_290_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_255_to_fp16 = const()[name = string("op_255_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30034176)))]; 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_255_to_fp16, x = var_290_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_256_to_fp16 = const()[name = string("op_256_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32131392)))]; 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_256_to_fp16, x = var_290_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_257_to_fp16 = const()[name = string("op_257_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32393600)))]; 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_257_to_fp16, x = var_290_cast_fp16_0)[name = string("value_states_7_cast_fp16")]; tensor concat_4x = const()[name = string("concat_4x"), val = tensor([-1, 16, 64, 5])]; 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, 5])]; 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, 5])]; 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_316_cast_fp16 = mul(x = embed_5_cast_fp16, y = cos_to_fp16)[name = string("op_316_cast_fp16")]; tensor var_317_split_sizes_0 = const()[name = string("op_317_split_sizes_0"), val = tensor([32, 32])]; int32 var_317_axis_0 = const()[name = string("op_317_axis_0"), val = int32(-2)]; 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 = embed_5_cast_fp16)[name = string("op_317_cast_fp16")]; fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_319_cast_fp16 = mul(x = var_317_cast_fp16_1, y = const_6_promoted_to_fp16)[name = string("op_319_cast_fp16")]; bool var_321_interleave_0 = const()[name = string("op_321_interleave_0"), val = bool(false)]; tensor var_321_cast_fp16 = concat(axis = var_260, interleave = var_321_interleave_0, values = (var_319_cast_fp16, var_317_cast_fp16_0))[name = string("op_321_cast_fp16")]; tensor var_322_cast_fp16 = mul(x = var_321_cast_fp16, y = sin_to_fp16)[name = string("op_322_cast_fp16")]; tensor query_states_9_cast_fp16 = add(x = var_316_cast_fp16, y = var_322_cast_fp16)[name = string("query_states_9_cast_fp16")]; tensor var_324_cast_fp16 = mul(x = embed_7_cast_fp16, y = cos_to_fp16)[name = string("op_324_cast_fp16")]; tensor var_325_split_sizes_0 = const()[name = string("op_325_split_sizes_0"), val = tensor([32, 32])]; int32 var_325_axis_0 = const()[name = string("op_325_axis_0"), val = int32(-2)]; tensor var_325_cast_fp16_0, tensor var_325_cast_fp16_1 = split(axis = var_325_axis_0, split_sizes = var_325_split_sizes_0, x = embed_7_cast_fp16)[name = string("op_325_cast_fp16")]; fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_327_cast_fp16 = mul(x = var_325_cast_fp16_1, y = const_7_promoted_to_fp16)[name = string("op_327_cast_fp16")]; bool var_329_interleave_0 = const()[name = string("op_329_interleave_0"), val = bool(false)]; tensor var_329_cast_fp16 = concat(axis = var_260, interleave = var_329_interleave_0, values = (var_327_cast_fp16, var_325_cast_fp16_0))[name = string("op_329_cast_fp16")]; tensor var_330_cast_fp16 = mul(x = var_329_cast_fp16, y = sin_to_fp16)[name = string("op_330_cast_fp16")]; tensor key_states_9_cast_fp16 = add(x = var_324_cast_fp16, y = var_330_cast_fp16)[name = string("key_states_9_cast_fp16")]; tensor var_335_split_sizes_0 = const()[name = string("op_335_split_sizes_0"), val = tensor([8, 8])]; int32 var_335_axis_0 = const()[name = string("op_335_axis_0"), val = int32(1)]; tensor var_335_cast_fp16_0, tensor var_335_cast_fp16_1 = split(axis = var_335_axis_0, split_sizes = var_335_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_335_cast_fp16")]; tensor var_337_split_sizes_0 = const()[name = string("op_337_split_sizes_0"), val = tensor([1, 1])]; int32 var_337_axis_0 = const()[name = string("op_337_axis_0"), val = int32(1)]; tensor var_337_cast_fp16_0, tensor var_337_cast_fp16_1 = split(axis = var_337_axis_0, split_sizes = var_337_split_sizes_0, x = key_states_9_cast_fp16)[name = string("op_337_cast_fp16")]; tensor var_339_split_sizes_0 = const()[name = string("op_339_split_sizes_0"), val = tensor([1, 1])]; int32 var_339_axis_0 = const()[name = string("op_339_axis_0"), val = int32(1)]; tensor var_339_cast_fp16_0, tensor var_339_cast_fp16_1 = split(axis = var_339_axis_0, split_sizes = var_339_split_sizes_0, x = value_states_9_cast_fp16)[name = string("op_339_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_337_cast_fp16_0, y = var_335_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_269, x = _inversed_attn_weights_15_cast_fp16)[name = string("attn_weights_17_cast_fp16")]; bool var_346_transpose_x_0 = const()[name = string("op_346_transpose_x_0"), val = bool(false)]; bool var_346_transpose_y_0 = const()[name = string("op_346_transpose_y_0"), val = bool(false)]; tensor var_346_cast_fp16 = matmul(transpose_x = var_346_transpose_x_0, transpose_y = var_346_transpose_y_0, x = var_339_cast_fp16_0, y = attn_weights_17_cast_fp16)[name = string("op_346_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_337_cast_fp16_1, y = var_335_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_269, 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_339_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_264, interleave = attn_output_7_interleave_0, values = (var_346_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, 5])]; 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_263_to_fp16 = const()[name = string("op_263_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32655808)))]; 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_263_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_365_cast_fp16 = mul(x = x_31_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_365_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_264, interleave = x_33_interleave_0, values = (x_31_cast_fp16, var_365_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_375_to_fp16 = const()[name = string("op_375_to_fp16"), val = fp16(0x1.5p-17)]; tensor out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_375_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(34753024)))]; 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_381_split_sizes_0 = const()[name = string("op_381_split_sizes_0"), val = tensor([1024, 1024])]; int32 var_381_axis_0 = const()[name = string("op_381_axis_0"), val = int32(1)]; tensor var_381_cast_fp16_0, tensor var_381_cast_fp16_1 = split(axis = var_381_axis_0, split_sizes = var_381_split_sizes_0, x = out_21_cast_fp16)[name = string("op_381_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_250_to_fp16 = const()[name = string("op_250_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34757184)))]; 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_250_to_fp16, x = var_381_cast_fp16_0)[name = string("input_3_cast_fp16")]; tensor var_389_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_389_cast_fp16")]; string var_394_pad_type_0 = const()[name = string("op_394_pad_type_0"), val = string("valid")]; tensor var_394_strides_0 = const()[name = string("op_394_strides_0"), val = tensor([1, 1])]; tensor var_394_pad_0 = const()[name = string("op_394_pad_0"), val = tensor([0, 0, 0, 0])]; tensor var_394_dilations_0 = const()[name = string("op_394_dilations_0"), val = tensor([1, 1])]; int32 var_394_groups_0 = const()[name = string("op_394_groups_0"), val = int32(1)]; tensor var_251_to_fp16 = const()[name = string("op_251_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43145856)))]; tensor var_394_cast_fp16 = conv(dilations = var_394_dilations_0, groups = var_394_groups_0, pad = var_394_pad_0, pad_type = var_394_pad_type_0, strides = var_394_strides_0, weight = var_251_to_fp16, x = var_381_cast_fp16_0)[name = string("op_394_cast_fp16")]; tensor x_39_cast_fp16 = mul(x = var_389_cast_fp16, y = var_394_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_252_to_fp16 = const()[name = string("op_252_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51534528)))]; 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_252_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_412 = const()[name = string("op_412"), val = int32(-2)]; int32 var_416 = const()[name = string("op_416"), val = int32(1)]; int32 var_421 = const()[name = string("op_421"), val = int32(2)]; fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_426_cast_fp16 = mul(x = x_41_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_426_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_416, interleave = x_43_interleave_0, values = (x_41_cast_fp16, var_426_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_436_to_fp16 = const()[name = string("op_436_to_fp16"), val = fp16(0x1.5p-17)]; tensor out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_436_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(59923200)))]; 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_442_split_sizes_0 = const()[name = string("op_442_split_sizes_0"), val = tensor([1024, 1024])]; int32 var_442_axis_0 = const()[name = string("op_442_axis_0"), val = int32(1)]; tensor var_442_cast_fp16_0, tensor var_442_cast_fp16_1 = split(axis = var_442_axis_0, split_sizes = var_442_split_sizes_0, x = out_27_cast_fp16)[name = string("op_442_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_407_to_fp16 = const()[name = string("op_407_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59927360)))]; 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_407_to_fp16, x = var_442_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_408_to_fp16 = const()[name = string("op_408_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62024576)))]; 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_408_to_fp16, x = var_442_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_409_to_fp16 = const()[name = string("op_409_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62286784)))]; 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_409_to_fp16, x = var_442_cast_fp16_0)[name = string("value_states_13_cast_fp16")]; tensor concat_8x = const()[name = string("concat_8x"), val = tensor([-1, 16, 64, 5])]; 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, 5])]; 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, 5])]; 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_468_cast_fp16 = mul(x = embed_9_cast_fp16, y = cos_to_fp16)[name = string("op_468_cast_fp16")]; tensor var_469_split_sizes_0 = const()[name = string("op_469_split_sizes_0"), val = tensor([32, 32])]; int32 var_469_axis_0 = const()[name = string("op_469_axis_0"), val = int32(-2)]; 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 = embed_9_cast_fp16)[name = string("op_469_cast_fp16")]; fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_471_cast_fp16 = mul(x = var_469_cast_fp16_1, y = const_10_promoted_to_fp16)[name = string("op_471_cast_fp16")]; bool var_473_interleave_0 = const()[name = string("op_473_interleave_0"), val = bool(false)]; tensor var_473_cast_fp16 = concat(axis = var_412, interleave = var_473_interleave_0, values = (var_471_cast_fp16, var_469_cast_fp16_0))[name = string("op_473_cast_fp16")]; tensor var_474_cast_fp16 = mul(x = var_473_cast_fp16, y = sin_to_fp16)[name = string("op_474_cast_fp16")]; tensor query_states_15_cast_fp16 = add(x = var_468_cast_fp16, y = var_474_cast_fp16)[name = string("query_states_15_cast_fp16")]; tensor var_476_cast_fp16 = mul(x = embed_11_cast_fp16, y = cos_to_fp16)[name = string("op_476_cast_fp16")]; tensor var_477_split_sizes_0 = const()[name = string("op_477_split_sizes_0"), val = tensor([32, 32])]; int32 var_477_axis_0 = const()[name = string("op_477_axis_0"), val = int32(-2)]; tensor var_477_cast_fp16_0, tensor var_477_cast_fp16_1 = split(axis = var_477_axis_0, split_sizes = var_477_split_sizes_0, x = embed_11_cast_fp16)[name = string("op_477_cast_fp16")]; fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)]; tensor var_479_cast_fp16 = mul(x = var_477_cast_fp16_1, y = const_11_promoted_to_fp16)[name = string("op_479_cast_fp16")]; bool var_481_interleave_0 = const()[name = string("op_481_interleave_0"), val = bool(false)]; tensor var_481_cast_fp16 = concat(axis = var_412, interleave = var_481_interleave_0, values = (var_479_cast_fp16, var_477_cast_fp16_0))[name = string("op_481_cast_fp16")]; tensor var_482_cast_fp16 = mul(x = var_481_cast_fp16, y = sin_to_fp16)[name = string("op_482_cast_fp16")]; tensor key_states_15_cast_fp16 = add(x = var_476_cast_fp16, y = var_482_cast_fp16)[name = string("key_states_15_cast_fp16")]; tensor var_487_split_sizes_0 = const()[name = string("op_487_split_sizes_0"), val = tensor([8, 8])]; int32 var_487_axis_0 = const()[name = string("op_487_axis_0"), val = int32(1)]; tensor var_487_cast_fp16_0, tensor var_487_cast_fp16_1 = split(axis = var_487_axis_0, split_sizes = var_487_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_487_cast_fp16")]; tensor var_489_split_sizes_0 = const()[name = string("op_489_split_sizes_0"), val = tensor([1, 1])]; int32 var_489_axis_0 = const()[name = string("op_489_axis_0"), val = int32(1)]; tensor var_489_cast_fp16_0, tensor var_489_cast_fp16_1 = split(axis = var_489_axis_0, split_sizes = var_489_split_sizes_0, x = key_states_15_cast_fp16)[name = string("op_489_cast_fp16")]; tensor var_491_split_sizes_0 = const()[name = string("op_491_split_sizes_0"), val = tensor([1, 1])]; int32 var_491_axis_0 = const()[name = string("op_491_axis_0"), val = int32(1)]; tensor var_491_cast_fp16_0, tensor var_491_cast_fp16_1 = split(axis = var_491_axis_0, split_sizes = var_491_split_sizes_0, x = value_states_15_cast_fp16)[name = string("op_491_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_489_cast_fp16_0, y = var_487_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_421, x = _inversed_attn_weights_27_cast_fp16)[name = string("attn_weights_29_cast_fp16")]; bool var_498_transpose_x_0 = const()[name = string("op_498_transpose_x_0"), val = bool(false)]; bool var_498_transpose_y_0 = const()[name = string("op_498_transpose_y_0"), val = bool(false)]; tensor var_498_cast_fp16 = matmul(transpose_x = var_498_transpose_x_0, transpose_y = var_498_transpose_y_0, x = var_491_cast_fp16_0, y = attn_weights_29_cast_fp16)[name = string("op_498_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_489_cast_fp16_1, y = var_487_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