Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

reland D70126859 #2787

Closed
wants to merge 1 commit into from
Closed

Conversation

TroyGarden
Copy link
Contributor

Differential Revision: D70712348

@pytorch-bot pytorch-bot bot added the ci-no-td label Mar 7, 2025
@facebook-github-bot facebook-github-bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Mar 7, 2025
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D70712348

@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D70712348

@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D70712348

TroyGarden added a commit to TroyGarden/torchrec that referenced this pull request Mar 8, 2025
Summary:

# context
* previous diff triggered S495021 
* the error message is like
```
ModelGenerationPlatformError("AttributeError: '_EmbeddingBagProxy' object has no attribute 'weight'")
```
* this diff works around the `embedding_bag.weight` to access the weight.dtype, instead, using the dtype from the table config.

Differential Revision: D70712348
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D70712348

Summary:

# context
* previous diff triggered S495021 
* the error message is like
```
ModelGenerationPlatformError("AttributeError: '_EmbeddingBagProxy' object has no attribute 'weight'")
```
* This is because in some flow the EBC module is fx traced so there is no actual EBC but a Proxy. Without full context it's risky to push this change.
* as a workaround, we'll just convert the unsharded EBC back to float32 so it's compatible with the input KJT.weight of float32

NOTE: this hacky change (unsharded EBC float16 ==> float32) is only needed in the tests, where we want to compare the results from sharded EBC. 

WARNING: We make a strong assumption here that in any unsharded EBC (with dtype=float16) use case, the input KJT.weights should never be float32.

Reviewed By: basilwong

Differential Revision: D70712348
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D70712348

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
ci-no-td CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. fb-exported
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants