Split bf16 and fp16 out for CK fp8_rowwise #4419
Open
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify project configuration. |
|
This pull request was exported from Phabricator. Differential Revision: D77544503 |
Summary: Pull Request resolved: pytorch#4419 While integrating AMD fp8 rowwise in torch, I noticed the library size is a lot larger than FBGEMM 1.2.0. The cause is we recently added support for fp16 output (in addition to bf16) in fp8 rowwise in D74770197, so this would cause the size to double. You can also see this in the nightly wheel. This diff does 2 changes: - Split out the fp16 out into their own files, while sharing common CK kernel implementation template `fp8_rowwise_common.h`, this will let us control which kernels to add in, e.g. only add bf16 out for torch. - For bf16 we will use `fp8fp8bf16` filename prefix - For fp16 we will use `fp8fp8fp16` filename prefix - The fp16 output is for a future rec-sys related use-case, so we can probably remove the llama specific tuning/optimized kernels for now (also reducing the size for that kernel). We can add new tuning to the fp16 out use-case for recsys after. Differential Revision: D77544503
|
This pull request was exported from Phabricator. Differential Revision: D77544503 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary:
While integrating AMD fp8 rowwise in torch, I noticed the library size is a lot larger than FBGEMM 1.2.0. The cause is we recently added support for fp16 output (in addition to bf16) in fp8 rowwise in D74770197, so this would cause the size to double. You can also see this in the nightly wheel - AMD goes from
This diff does 2 changes:
fp8_rowwise_common.h, this will let us control which kernels to add in, e.g. only add bf16 out for torch.Differential Revision: D77544503