Issue #4635: Use ml_dtypes to fix onnx_graphsurgeon import error #4653
Conversation
onnx 1.20 has removed some float convcersion helpers in favor of using ml_dtypes package. Use ml_dtypes to fix onnx_graphsurgeon import errors. Fixes NVIDIA#4635 Signed-off-by: Manoj Gupta <[email protected]>
m-gupta
changed the title
|
@kevinch-nv @yuanyao-nv friendly ping for PR that would help us to upgrade newest onnx as well. Thank you in advance !! :) |
yuanyao-nv
left a comment
yuanyao-nv
left a comment
There was a problem hiding this comment.
Thanks for the contributions!
|
Hello again, @yuanyao-nv thank you for merging, great to see it. Any time line for release to "pypi" ? That would great great to know. Thank you in advance |
|
@onuralpszr Have you tried running the tests. I noticed some tests are broken by these changes. |
|
Patch author here. Are there any instructions on how to run the tests? And shouldn't CI have caught any issues. |
|
@m-gupta You can just run pytest in the graph-surgeon directory. Unfortunately this repo doesn't currently run CI as it's not a development repo. |
|
@yuanyao-nv I currently do not have time to look into it. Despite what ml_dtypes say, the values are not really same atleast when printing and comparing though tobytes return the same thing. Also numpy float32 does not account for shape. Sp Please feel free to revert. |
3 participants
onnx 1.20 has removed some float convcersion helpers in favor of using ml_dtype package. Use ml_dtypes to fix onnx_graphsurgeon import errors.
Fixes #4635