fix fp8 kv cache dequantize kernels #3896
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: Fix fp8 kv cache dequantization kernel and enable unit test on AMD. The kernel uses each thread to dequantize 4 elements for both K and V and each warp for a head. The dim is always 128. So on NV this works as one warp has 32 threads on NV (4 * 32 = 128). On AMD, each wavefront (warp) has 64 threads, so the second 32 threads will all do out-of-bound memory access.... This diff simply masks those threads to do nothing. Obviously the perf is not good but from E2E testing, it does not seem to matter. If we need to optimize the perf for AMD, we can let thread 0 ~ 31 dequantize 4 elements for K and thread 32 ~ 63 thread dequantize 4 elements for V. Differential Revision: D72062745
|
This pull request was exported from Phabricator. Differential Revision: D72062745 |
✅ 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 a303797. |
|
This pull request has been reverted by 47635cf. |
q10
pushed a commit
to q10/FBGEMM
that referenced
this pull request
Apr 10, 2025
Summary: X-link: pytorch#3896 Pull Request resolved: facebookresearch/FBGEMM#987 Fix fp8 kv cache dequantization kernel and enable unit test on AMD. The kernel uses each thread to dequantize 4 elements for both K and V and each warp for a head. The dim is always 128. So on NV this works as one warp has 32 threads on NV (4 * 32 = 128). On AMD, each wavefront (warp) has 64 threads, so the second 32 threads will all do out-of-bound memory access.... This diff simply masks those threads to do nothing. Obviously the perf is not good but from E2E testing, it does not seem to matter. If we need to optimize the perf for AMD, we can let thread 0 ~ 31 dequantize 4 elements for K and thread 32 ~ 63 thread dequantize 4 elements for V. Reviewed By: Aya-ZIbra Differential Revision: D72062745 fbshipit-source-id: 1b813057586054a13df4e9088be00b08f912bc57
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:
Fix fp8 kv cache dequantization kernel and enable unit test on AMD.
The kernel uses each thread to dequantize 4 elements for both K and V and each warp for a head. The dim is always 128. So on NV this works as one warp has 32 threads on NV (4 * 32 = 128).
On AMD, each wavefront (warp) has 64 threads, so the second 32 threads will all do out-of-bound memory access....
This diff simply masks those threads to do nothing. Obviously the perf is not good but from E2E testing, it does not seem to matter. If we need to optimize the perf for AMD, we can let thread 0 ~ 31 dequantize 4 elements for K and thread 32 ~ 63 thread dequantize 4 elements for V.
Differential Revision: D72062745