Skip to content

Register sdp lower precision autocast#7299

Merged
JackCaoG merged 2 commits intomasterfrom
unknown repository
Jun 25, 2024
Merged

Register sdp lower precision autocast#7299
JackCaoG merged 2 commits intomasterfrom
unknown repository

Conversation

@ghost
Copy link
Copy Markdown

@ghost ghost commented Jun 17, 2024

Details can be seen in Issue 7177.

@ghost
Copy link
Copy Markdown
Author

ghost commented Jun 18, 2024

Most of the failure caused by numpy, seem no related about this PR. How can I rerun this PR after the fix of build/test enviroment?

@JackCaoG
Copy link
Copy Markdown
Collaborator

hmm if you rebase the issue should be gone

@ghost
Copy link
Copy Markdown
Author

ghost commented Jun 20, 2024

I think this PR need re-approve to run the tests.

@ghost
Copy link
Copy Markdown
Author

ghost commented Jun 21, 2024

I can not reproduce the test_train_mp_mnist_amp.py failure in my machine. And it is weird that the training loss has decreased to 0.00066 which is same as I see in my machine but the accuracy is 0. I really don’t know why sdp affects the mnist. It’s obvious that this op is not used at all.

@JackCaoG
Copy link
Copy Markdown
Collaborator

Let me retrigger, it might be using the fake data so accuracy doesn't matter. It seems like CI just exited with an error code but I don't know why.

@ghost
Copy link
Copy Markdown
Author

ghost commented Jun 25, 2024

The CI has passed, this PR can be merged.

@JackCaoG JackCaoG merged commit 53c77e2 into pytorch:master Jun 25, 2024
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.

3 participants