Felladrin commited on
Commit
af0ef30
·
1 Parent(s): e3feed5

Pin optimum export stack to reproducible versions

Browse files

The previous fix relied on pip happening to resolve a compatible
optimum/optimum-onnx/transformers combination on Python 3.13. With
those left unpinned, a resolver drift (or a future 3.14 base-image
bump) could reintroduce the NormalizedConfig 'allow_new' TypeError
that broke every conversion.

Pin the three packages that determine exporter behavior to the
versions verified working (optimum 2.1.0, optimum-onnx 0.1.0,
transformers 4.57.6), and pin onnx-ir 0.2.1 since onnxruntime imports
it lazily without declaring it. The onnx/onnxruntime/onnxconverter-common
floors are kept as-is; they encode real minimums (quantizer classes,
float16 API stability).

Files changed (1) hide show
  1. requirements.txt +4 -2
requirements.txt CHANGED
@@ -3,8 +3,10 @@ streamlit==1.50.0
3
  PyYAML==6.0.2
4
  torch==2.5.1; python_version < "3.13"
5
  torch>=2.9.0; python_version >= "3.13"
6
- optimum[onnx]
 
 
7
  onnx>=1.22
8
  onnxruntime>=1.27
9
- onnx-ir
10
  onnxconverter-common>=1.16,<1.17
 
3
  PyYAML==6.0.2
4
  torch==2.5.1; python_version < "3.13"
5
  torch>=2.9.0; python_version >= "3.13"
6
+ optimum[onnx]==2.1.0
7
+ optimum-onnx==0.1.0
8
+ transformers==4.57.6
9
  onnx>=1.22
10
  onnxruntime>=1.27
11
+ onnx-ir==0.2.1
12
  onnxconverter-common>=1.16,<1.17