-
Notifications
You must be signed in to change notification settings - Fork 2.3k
Open
Labels
Module:GraphSurgeonIssues with ONNX-GraphsurgeonIssues with ONNX-GraphsurgeonModule:ONNXIssues relating to ONNX usage and importIssues relating to ONNX usage and import
Description
onnx.helper dtypes have been deprecated in ONNX 1.20.0. So onnx_graphsurgeon import fails
import onnx_graphsurgeon
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
Cell In[1], line 1
----> 1 import onnx_graphsurgeon
File /opt/conda/lib/python3.11/site-packages/onnx_graphsurgeon/__init__.py:1
----> 1 from onnx_graphsurgeon.exporters.onnx_exporter import export_onnx
2 from onnx_graphsurgeon.graph_pattern import GraphPattern, PatternMapping
3 from onnx_graphsurgeon.importers.onnx_importer import import_onnx
File /opt/conda/lib/python3.11/site-packages/onnx_graphsurgeon/exporters/onnx_exporter.py:134
128 def __call__(self, arr):
129 return self.func(arr)
132 _NUMPY_ARRAY_CONVERTERS = {
133 onnx.TensorProto.BFLOAT16: NumpyArrayConverter(
--> 134 np.uint16, onnx.helper.float32_to_bfloat16
135 ),
136 # FP8 in TensorRT supports negative zeros, no infinities
137 # See https://onnx.ai/onnx/technical/float8.html#papers
138 onnx.TensorProto.FLOAT8E4M3FN: NumpyArrayConverter(
139 np.uint8, lambda x: onnx.helper.float32_to_float8e4m3(x, fn=True, uz=False)
140 ),
141 }
144 # Converts a gs.Constant to an onnx tensor and convert the values according to gs.Constant.export_dtype
145 def constant_to_onnx_tensor(tensor: Constant) -> onnx.TensorProto:
AttributeError: module 'onnx.helper' has no attribute 'float32_to_bfloat16'https://github.com/onnx/onnx/blob/8827114d2c90c924ab05b68ac1093b0a6adf9324/onnx/helper.py#L359
Reproduce
pip install --pre -U onnximport onnx_graphsurgeon
gredin
Metadata
Metadata
Assignees
Labels
Module:GraphSurgeonIssues with ONNX-GraphsurgeonIssues with ONNX-GraphsurgeonModule:ONNXIssues relating to ONNX usage and importIssues relating to ONNX usage and import