Skip to content

Fix IOBinding population error when batch_size>1#461

Merged
JingyaHuang merged 2 commits into
mainfrom
fix-io-binding
Nov 10, 2022
Merged

Fix IOBinding population error when batch_size>1#461
JingyaHuang merged 2 commits into
mainfrom
fix-io-binding

Conversation

@JingyaHuang

Copy link
Copy Markdown
Contributor

What does this PR do?

Fixes #404

  • Force tensors to be contiguous (otherwise might populate in the wrong address).
  • Add bs>1 and beam search to test.

Kudos to @soocheolnoh for testing and catching the bug.

@HuggingFaceDocBuilderDev

HuggingFaceDocBuilderDev commented Nov 10, 2022

Copy link
Copy Markdown

The documentation is not available anymore as the PR was closed or merged.

@fxmarty

fxmarty commented Nov 10, 2022

Copy link
Copy Markdown
Contributor

LGTM. Is this fixing the issue of translation results being different? If there is no such test, can we add one?

@JingyaHuang

Copy link
Copy Markdown
Contributor Author

LGTM. Is this fixing the issue of translation results being different? If there is no such test, can we add one?

Yes, it does. And I added beam search on the seq2seq generation test. I think it would be enough.

@mht-sharma

Copy link
Copy Markdown
Contributor

Thanks @JingyaHuang for the quick fix. LGTM too.

@JingyaHuang
JingyaHuang merged commit df66bd3 into main Nov 10, 2022
@JingyaHuang
JingyaHuang deleted the fix-io-binding branch November 10, 2022 13:55
@JingyaHuang
JingyaHuang restored the fix-io-binding branch November 24, 2022 10:27
@JingyaHuang
JingyaHuang deleted the fix-io-binding branch November 28, 2022 22:19
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Inference worse with onnxruntime-gpu than native pytorch for seq2seq model

4 participants