PhiΒΆ
<<<
import numpy as np
import torch
from transformers import PhiConfig
from transformers.models.phi.modeling_phi import PhiModel
from experimental_experiment.helpers import pretty_onnx
from experimental_experiment.torch_interpreter import to_onnx, ExportOptions
from onnx_diagnostic.torch_export_patches import torch_export_patches
def ids_tensor(shape, vocab_size):
total_dims = 1
for dim in shape:
total_dims *= dim
values = []
for _ in range(total_dims):
values.append(np.random.randint(0, vocab_size - 1))
return torch.tensor(data=values, dtype=torch.long).view(shape).contiguous()
config = PhiConfig(
hidden_size=32,
num_hidden_layers=2,
vocab_size=1024,
intermediate_size=16,
max_position_embeddings=512,
num_attention_heads=2,
num_key_value_heads=2,
)
config._attn_implementation = "eager"
with torch.no_grad(), torch_export_patches(patch_transformers=True) as modificator:
model = PhiModel(config)
batch, seq, vocab_size = 2, 1024, 1024
input_ids = ids_tensor([batch, seq], vocab_size)
input_mask = torch.tril(torch.ones(batch, seq, dtype=torch.float32))
model(input_ids, input_mask)
onx = to_onnx(
model,
modificator((input_ids, input_mask)),
export_options=ExportOptions(decomposition_table="default"),
)
print(pretty_onnx(onx))
>>>
opset: domain='' version=18
opset: domain='intermediate' version=1
input: name='input_ids' type=dtype('int64') shape=[2, 1024]
input: name='attention_mask' type=dtype('float32') shape=[2, 1024]
init: name='init7_s1_1' type=int64 shape=(1,) -- array([1]) -- Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape
init: name='init7_s1_-1' type=int64 shape=(1,) -- array([-1]) -- Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape
init: name='init7_s4_2_1024_-1_16' type=int64 shape=(4,) -- array([ 2, 1024, -1, 16])-- Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape
init: name='init7_s3_2_1024_-1' type=int64 shape=(3,) -- array([ 2, 1024, -1])-- Opset.make_node.1/Shape##Opset.make_node.1/Shape
init: name='init1_s_5' type=float32 shape=() -- array([3.], dtype=float32)-- Opset.make_node.1/Small##Opset.make_node.1/Small
init: name='and_1' type=bool shape=(2, 1, 1024, 1024) -- GraphBuilder.constant_folding.from/fold(_onx_and_new_ones::RSh1)##_onx_and_new_ones::RSh1/
init: name='_onx_add_expand_4' type=int64 shape=(2, 1, 1024, 1024) -- GraphBuilder.constant_folding.from/fold(_onx_mul_expand,expand_4)##expand_4/##_onx_mul_expand/GraphBuilder.constant_folding.from/fold(_to_copy::Shape-1:,expand)##expand/##_to_copy::Shape-1:/
init: name='_onx_add_expand_4::RSh-1' type=int64 shape=(2097152,) -- GraphBuilder.constant_folding.from/fold(_onx_add_expand_4,init7_s1_-1)##_onx_add_expand_4/GraphBuilder.constant_folding.from/fold(_onx_mul_expand,expand_4)##expand_4/##_onx_mul_expand/GraphBuilder.constant_folding.from/fold(_to_copy::Shape-1:,expand)##expand/##_to_copy::Shape-1:/##init7_s1_-1/Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape
init: name='init1_s_::RSh1' type=float32 shape=(1,) -- array([-3.403e+38], dtype=float32)-- GraphBuilder.constant_folding.from/fold(init1_s_,init7_s1_1)##init1_s_/shape_type_compute._cast_inputs.1(mul_Tensor)##init7_s1_1/Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape
init: name='p_layers_0_self_attn_q_proj_weight::T10' type=float32 shape=(32, 32)-- GraphBuilder.constant_folding.from/fold(p_layers_0_self_attn_q_proj_weight)##p_layers_0_self_attn_q_proj_weight/DynamoInterpret.placeholder.1/P(layers.0.self_attn.q_proj.weight)
init: name='p_layers_0_self_attn_k_proj_weight::T10' type=float32 shape=(32, 32)-- GraphBuilder.constant_folding.from/fold(p_layers_0_self_attn_k_proj_weight)##p_layers_0_self_attn_k_proj_weight/DynamoInterpret.placeholder.1/P(layers.0.self_attn.k_proj.weight)
init: name='p_layers_0_self_attn_v_proj_weight::T10' type=float32 shape=(32, 32)-- GraphBuilder.constant_folding.from/fold(p_layers_0_self_attn_v_proj_weight)##p_layers_0_self_attn_v_proj_weight/DynamoInterpret.placeholder.1/P(layers.0.self_attn.v_proj.weight)
init: name='init1_s_3::RSh1' type=float32 shape=(1,) -- array([0.25], dtype=float32)-- GraphBuilder.constant_folding.from/fold(init1_s_3,init7_s1_1)##init1_s_3/shape_type_compute._cast_inputs.1(mul_Tensor)##shape_type_compute._cast_inputs.1(mul_Tensor)##init7_s1_1/Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape
init: name='p_layers_0_self_attn_dense_weight::T10' type=float32 shape=(32, 32)-- GraphBuilder.constant_folding.from/fold(p_layers_0_self_attn_dense_weight)##p_layers_0_self_attn_dense_weight/DynamoInterpret.placeholder.1/P(layers.0.self_attn.dense.weight)
init: name='p_layers_0_mlp_fc1_weight::T10' type=float32 shape=(32, 16)-- GraphBuilder.constant_folding.from/fold(p_layers_0_mlp_fc1_weight)##p_layers_0_mlp_fc1_weight/DynamoInterpret.placeholder.1/P(layers.0.mlp.fc1.weight)
init: name='init1_s_4::RSh1' type=float32 shape=(1,) -- array([0.5], dtype=float32)-- GraphBuilder.constant_folding.from/fold(init1_s_4,init7_s1_1)##init1_s_4/shape_type_compute._cast_inputs.1(mul_Tensor)##shape_type_compute._cast_inputs.1(mul_Tensor)##init7_s1_1/Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape
init: name='init1_s_6::RSh1' type=float32 shape=(1,) -- array([0.045], dtype=float32)-- GraphBuilder.constant_folding.from/fold(init1_s_6,init7_s1_1)##init1_s_6/shape_type_compute._cast_inputs.1(mul_Tensor)##shape_type_compute._cast_inputs.1(mul_Tensor)##init7_s1_1/Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape
init: name='init1_s_7::RSh1' type=float32 shape=(1,) -- array([0.798], dtype=float32)-- GraphBuilder.constant_folding.from/fold(init1_s_7,init7_s1_1)##init1_s_7/shape_type_compute._cast_inputs.1(mul_Tensor)##shape_type_compute._cast_inputs.1(mul_Tensor)##init7_s1_1/Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape
init: name='init1_s_2::RSh13' type=float32 shape=(1,) -- array([1.], dtype=float32)-- GraphBuilder.constant_folding.from/fold(init1_s_2,init7_s1_1)##init1_s_2/shape_type_compute._cast_inputs.1(mul_Tensor)##shape_type_compute._cast_inputs.1(mul_Tensor)##shape_type_compute._cast_inputs.0##shape_type_compute._cast_inputs.0##init7_s1_1/Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape##Opset.make_node.1/Shape
init: name='p_layers_0_mlp_fc2_weight::T10' type=float32 shape=(16, 32)-- GraphBuilder.constant_folding.from/fold(p_layers_0_mlp_fc2_weight)##p_layers_0_mlp_fc2_weight/DynamoInterpret.placeholder.1/P(layers.0.mlp.fc2.weight)
init: name='p_layers_1_self_attn_q_proj_weight::T10' type=float32 shape=(32, 32)-- GraphBuilder.constant_folding.from/fold(p_layers_1_self_attn_q_proj_weight)##p_layers_1_self_attn_q_proj_weight/DynamoInterpret.placeholder.1/P(layers.1.self_attn.q_proj.weight)
init: name='p_layers_1_self_attn_k_proj_weight::T10' type=float32 shape=(32, 32)-- GraphBuilder.constant_folding.from/fold(p_layers_1_self_attn_k_proj_weight)##p_layers_1_self_attn_k_proj_weight/DynamoInterpret.placeholder.1/P(layers.1.self_attn.k_proj.weight)
init: name='p_layers_1_self_attn_v_proj_weight::T10' type=float32 shape=(32, 32)-- GraphBuilder.constant_folding.from/fold(p_layers_1_self_attn_v_proj_weight)##p_layers_1_self_attn_v_proj_weight/DynamoInterpret.placeholder.1/P(layers.1.self_attn.v_proj.weight)
init: name='p_layers_1_self_attn_dense_weight::T10' type=float32 shape=(32, 32)-- GraphBuilder.constant_folding.from/fold(p_layers_1_self_attn_dense_weight)##p_layers_1_self_attn_dense_weight/DynamoInterpret.placeholder.1/P(layers.1.self_attn.dense.weight)
init: name='p_layers_1_mlp_fc1_weight::T10' type=float32 shape=(32, 16)-- GraphBuilder.constant_folding.from/fold(p_layers_1_mlp_fc1_weight)##p_layers_1_mlp_fc1_weight/DynamoInterpret.placeholder.1/P(layers.1.mlp.fc1.weight)
init: name='p_layers_1_mlp_fc2_weight::T10' type=float32 shape=(16, 32)-- GraphBuilder.constant_folding.from/fold(p_layers_1_mlp_fc2_weight)##p_layers_1_mlp_fc2_weight/DynamoInterpret.placeholder.1/P(layers.1.mlp.fc2.weight)
init: name='init1_s32_' type=float32 shape=(32,) -- LayerNormalizationPattern.apply.scale##LayerNormalizationPattern.apply.scale##LayerNormalizationPattern.apply.scale
init: name='init1_s32_2' type=float32 shape=(32,) -- LayerNormalizationPattern.apply.bias##LayerNormalizationPattern.apply.bias##LayerNormalizationPattern.apply.bias
init: name='init7_s2_8_8' type=int64 shape=(2,) -- array([8, 8]) -- SlicesSplitPattern.apply.splits##SlicesSplitPattern.apply.splits##SlicesSplitPattern.apply.splits##SlicesSplitPattern.apply.splits
init: name='transpose' type=float32 shape=(1, 1024, 4) -- GraphBuilder.constant_folding.from/fold(_to_copy_22,unsqueeze_22)##unsqueeze_22/##_to_copy_22/
init: name='embed_tokens.weight' type=float32 shape=(1024, 32) -- DynamoInterpret.placeholder.1/P(embed_tokens.weight)
Cast(attention_mask, to=9) -> _to_copy
Reshape(_to_copy, init7_s1_-1) -> _to_copy::RSh-1
Gather(_to_copy::RSh-1, _onx_add_expand_4::RSh-1) -> _onx_gather__to_copy::RSh-1
Gather(embed_tokens.weight, input_ids) -> embedding
LayerNormalization(embedding, init1_s32_, init1_s32_2, axis=-1, epsilon=0.00, stash_type=1) -> _onx_div_sub_clone
MatMul(_onx_div_sub_clone, p_layers_0_self_attn_q_proj_weight::T10) -> _onx_matmul_layer_norm
Reshape(_onx_matmul_layer_norm, init7_s4_2_1024_-1_16) -> view_5
Transpose(view_5, perm=[0,2,1,3]) -> transpose_1
Split(transpose_1, init7_s2_8_8, axis=3) -> slice_1, slice_2
Shape(_onx_add_expand_4) -> _onx_add_expand_4::Shape:
Reshape(_onx_gather__to_copy::RSh-1, _onx_add_expand_4::Shape:) -> index
And(and_1, index) -> and_2
Not(and_2) -> bitwise_not
Cast(bitwise_not, to=1) -> _to_copy_1
Mul(_to_copy_1, init1_s_::RSh1) -> _onx_mul__to_copy_1
Sin(transpose) -> usin
Unsqueeze(usin, init7_s1_1) -> uunsqueeze_5
Concat(uunsqueeze_5, uunsqueeze_5, axis=-1) -> unsqueeze_5
Cos(transpose) -> ucos
Unsqueeze(ucos, init7_s1_1) -> uunsqueeze_4
Concat(uunsqueeze_4, uunsqueeze_4, axis=-1) -> unsqueeze_4
HalfRotaryEmbedding[intermediate](slice_1, unsqueeze_4, unsqueeze_5) -> add_1
Concat(add_1, slice_2, axis=-1) -> cat_3
MatMul(_onx_div_sub_clone, p_layers_0_self_attn_k_proj_weight::T10) -> _onx_matmul_layer_norm2
Reshape(_onx_matmul_layer_norm2, init7_s4_2_1024_-1_16) -> view_6
Transpose(view_6, perm=[0,2,1,3]) -> transpose_2
Split(transpose_2, init7_s2_8_8, axis=3) -> slice_3, slice_4
HalfRotaryEmbedding[intermediate](slice_3, unsqueeze_4, unsqueeze_5) -> add_2
Concat(add_2, slice_4, axis=-1) -> output_1
Transpose(output_1, perm=[0,1,3,2]) -> transpose_4
MatMul(cat_3, transpose_4) -> matmul_1
Mul(matmul_1, init1_s_3::RSh1) -> _onx_mul_matmul_1
Add(_onx_mul_matmul_1, _onx_mul__to_copy_1) -> add_3
Softmax(add_3, axis=-1) -> softmax
MatMul(_onx_div_sub_clone, p_layers_0_self_attn_v_proj_weight::T10) -> _onx_matmul_layer_norm3
Reshape(_onx_matmul_layer_norm3, init7_s4_2_1024_-1_16) -> view_7
Transpose(view_7, perm=[0,2,1,3]) -> output_3
MatMul(softmax, output_3) -> matmul_2
Transpose(matmul_2, perm=[0,2,1,3]) -> transpose_5
Reshape(transpose_5, init7_s3_2_1024_-1) -> view_8
MatMul(view_8, p_layers_0_self_attn_dense_weight::T10) -> _onx_matmul_view_8
MatMul(_onx_div_sub_clone, p_layers_0_mlp_fc1_weight::T10) -> _onx_matmul_layer_norm4
Mul(_onx_matmul_layer_norm4, init1_s_4::RSh1) -> _onx_mul_linear_4
Pow(_onx_matmul_layer_norm4, init1_s_5) -> pow_1
Mul(pow_1, init1_s_6::RSh1) -> _onx_mul_pow_1
Add(_onx_matmul_layer_norm4, _onx_mul_pow_1) -> add_4
Mul(add_4, init1_s_7::RSh1) -> _onx_mul_add_4
Tanh(_onx_mul_add_4) -> tanh
Add(tanh, init1_s_2::RSh13) -> add_5
Mul(_onx_mul_linear_4, add_5) -> mul_11
MatMul(mul_11, p_layers_0_mlp_fc2_weight::T10) -> _onx_matmul_mul_11
Add(_onx_matmul_view_8, _onx_matmul_mul_11) -> add_6
Add(add_6, embedding) -> add_7
LayerNormalization(add_7, init1_s32_, init1_s32_2, axis=-1, epsilon=0.00, stash_type=1) -> _onx_div_sub_add_7
MatMul(_onx_div_sub_add_7, p_layers_1_self_attn_q_proj_weight::T10) -> _onx_matmul_layer_norm_1
Reshape(_onx_matmul_layer_norm_1, init7_s4_2_1024_-1_16) -> view_9
Transpose(view_9, perm=[0,2,1,3]) -> transpose_6
Split(transpose_6, init7_s2_8_8, axis=3) -> slice_9, slice_10
HalfRotaryEmbedding[intermediate](slice_9, unsqueeze_4, unsqueeze_5) -> add_8
Concat(add_8, slice_10, axis=-1) -> cat_9
MatMul(_onx_div_sub_add_7, p_layers_1_self_attn_k_proj_weight::T10) -> _onx_matmul_layer_norm_12
Reshape(_onx_matmul_layer_norm_12, init7_s4_2_1024_-1_16) -> view_10
Transpose(view_10, perm=[0,2,1,3]) -> transpose_7
Split(transpose_7, init7_s2_8_8, axis=3) -> slice_11, slice_12
HalfRotaryEmbedding[intermediate](slice_11, unsqueeze_4, unsqueeze_5) -> add_9
Concat(add_9, slice_12, axis=-1) -> output_2
Transpose(output_2, perm=[0,1,3,2]) -> transpose_9
MatMul(cat_9, transpose_9) -> matmul_3
Mul(matmul_3, init1_s_3::RSh1) -> _onx_mul_matmul_3
Add(_onx_mul_matmul_3, _onx_mul__to_copy_1) -> add_10
Softmax(add_10, axis=-1) -> softmax_1
MatMul(_onx_div_sub_add_7, p_layers_1_self_attn_v_proj_weight::T10) -> _onx_matmul_layer_norm_13
Reshape(_onx_matmul_layer_norm_13, init7_s4_2_1024_-1_16) -> view_11
Transpose(view_11, perm=[0,2,1,3]) -> output_4
MatMul(softmax_1, output_4) -> matmul_4
Transpose(matmul_4, perm=[0,2,1,3]) -> transpose_10
Reshape(transpose_10, init7_s3_2_1024_-1) -> view_12
MatMul(view_12, p_layers_1_self_attn_dense_weight::T10) -> _onx_matmul_view_12
MatMul(_onx_div_sub_add_7, p_layers_1_mlp_fc1_weight::T10) -> _onx_matmul_layer_norm_14
Mul(_onx_matmul_layer_norm_14, init1_s_4::RSh1) -> _onx_mul_linear_10
Pow(_onx_matmul_layer_norm_14, init1_s_5) -> pow_2
Mul(pow_2, init1_s_6::RSh1) -> _onx_mul_pow_2
Add(_onx_matmul_layer_norm_14, _onx_mul_pow_2) -> add_11
Mul(add_11, init1_s_7::RSh1) -> _onx_mul_add_11
Tanh(_onx_mul_add_11) -> tanh_1
Add(tanh_1, init1_s_2::RSh13) -> add_12
Mul(_onx_mul_linear_10, add_12) -> mul_20
MatMul(mul_20, p_layers_1_mlp_fc2_weight::T10) -> _onx_matmul_mul_20
Add(_onx_matmul_view_12, _onx_matmul_mul_20) -> add_13
Add(add_13, add_7) -> add_14
LayerNormalization(add_14, init1_s32_, init1_s32_2, axis=-1, epsilon=0.00, stash_type=1) -> output_0
output: name='output_0' type=dtype('float32') shape=[2, 1024, 32]
output: name='output_1' type=dtype('float32') shape=[2, 2, 1024, 16]
output: name='output_2' type=dtype('float32') shape=[2, 2, 1024, 16]
output: name='output_3' type=dtype('float32') shape=[2, 2, 1024, 16]
output: name='output_4' type=dtype('float32') shape=[2, 2, 1024, 16]
----- function name=HalfRotaryEmbedding domain=intermediate
----- doc_string: -- function_options=FunctionOptions(export_as_function=...
opset: domain='' version=18
input: 'X'
input: 'cos_cache'
input: 'sin_cache'
Mul(X, cos_cache) -> _onx_mul_X
Split(X, axis=-1, num_outputs=2) -> _onx_split_X_0, _onx_split_X_1
Neg(_onx_split_X_1) -> _onx_neg_split_X_1
Concat(_onx_neg_split_X_1, _onx_split_X_0, axis=-1) -> _onx_concat_neg_split_X_1
Mul(_onx_concat_neg_split_X_1, sin_cache) -> _onx_mul_concat_neg_split_X_1
Add(_onx_mul_X, _onx_mul_concat_neg_split_X_1) -> Y
output: name='Y' type=? shape=?