Enable fast FP8 GEMM for memory bound (resubmit) #3608
Closed
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
Summary: This Diff (resubmit from D68193920) enables fast FP8 gemm for memory bound with adding TRT-LLM FP8 cuda gemm to fbgemm. In addition to the original kernel, this Diff extends the kernel to: - Support pytorch operations - Support cuda graph with handling scale as tensor - Support smaller dim M for much faster compilation time - Support benchmark/unittest For decode attn linear shapes: - When BS=1, TRT-LLM FP8 gemm brings 2x speedup compared to BF16, while FP8 cutlass gemm’s perf is similar to BF16 - When BS>4, TRT-LLM FP8 gemm does not bring perf gain - This TRT-LLM kernel is based on tensorwise quantization not rowwise. Note: As M>4 does not bring perf gain in our use cases, we only instantiate 4 template instances to reduce compilation time (10 mins -> 2.5 mins). If we would like to increase instances for larger M in the future, we could tradeoff acceptable compilation time or dedicate cuda file to each instance with compile in parallel Differential Revision: D68568596
|
This pull request was exported from Phabricator. Differential Revision: D68568596 |
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify site configuration. |
|
This pull request has been merged in 5754ce7. |
q10
pushed a commit
to q10/FBGEMM
that referenced
this pull request
Apr 10, 2025
Summary: X-link: pytorch#3608 Pull Request resolved: facebookresearch/FBGEMM#686 This Diff (resubmit from D68193920) enables fast FP8 gemm for memory bound with adding TRT-LLM FP8 cuda gemm to fbgemm. In addition to the original kernel, this Diff extends the kernel to: - Support pytorch operations - Support cuda graph with handling scale as tensor - Support smaller dim M for much faster compilation time - Support benchmark/unittest For decode attn linear shapes: - When BS=1, TRT-LLM FP8 gemm brings 2x speedup compared to BF16, while FP8 cutlass gemm’s perf is similar to BF16 - When BS>4, TRT-LLM FP8 gemm does not bring perf gain - This TRT-LLM kernel is based on tensorwise quantization not rowwise. Note: As M>4 does not bring perf gain in our use cases, we only instantiate 4 template instances to reduce compilation time (10 mins -> 2.5 mins). If we would like to increase instances for larger M in the future, we could tradeoff acceptable compilation time or dedicate cuda file to each instance with compile in parallel Reviewed By: q10, jwfromm Differential Revision: D68568596 fbshipit-source-id: ba8b565a564533717deb29f9d701550d99a8c759
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:
This Diff (resubmit from D68193920) enables fast FP8 gemm for memory bound with adding TRT-LLM FP8 cuda gemm to fbgemm. In addition to the original kernel, this Diff extends the kernel to:
For decode attn linear shapes:
Note: As M>4 does not bring perf gain in our use cases, we only instantiate 4 template instances to reduce compilation time (10 mins -> 2.5 mins). If we would like to increase instances for larger M in the future, we could tradeoff acceptable compilation time or dedicate cuda file to each instance with compile in parallel
Differential Revision: D68568596