Fuse cumulative sum into FP8xINT4 Grouped Gemm #3812
Conversation
This file contains 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
|
This pull request was exported from Phabricator. Differential Revision: D71081537 |
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify site configuration. |
jwfromm
force-pushed
the
export-D71081537
branch
from
7b14029 to
ae44862
Compare
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Mar 13, 2025
Summary: X-link: facebookresearch/FBGEMM#898 Rather than do a separate cumsum operator in torch (which can be expensive), its quite trivial to fuse the sum into the kernel setup we have to do for GEMM anyway. This diff makes that change and we see no measurable performance impact. Reviewed By: jiawenliu64 Differential Revision: D71081537
jwfromm
force-pushed
the
export-D71081537
branch
from
ae44862 to
9fcd7c3
Compare
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Mar 13, 2025
Summary: X-link: facebookresearch/FBGEMM#898 Rather than do a separate cumsum operator in torch (which can be expensive), its quite trivial to fuse the sum into the kernel setup we have to do for GEMM anyway. This diff makes that change and we see no measurable performance impact. Reviewed By: jiawenliu64 Differential Revision: D71081537
|
This pull request was exported from Phabricator. Differential Revision: D71081537 |
1 similar comment
|
This pull request was exported from Phabricator. Differential Revision: D71081537 |
jwfromm
force-pushed
the
export-D71081537
branch
from
9fcd7c3 to
1bf1da4
Compare
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Mar 13, 2025
Summary: X-link: facebookresearch/FBGEMM#898 Pull Request resolved: pytorch#3812 Rather than do a separate cumsum operator in torch (which can be expensive), its quite trivial to fuse the sum into the kernel setup we have to do for GEMM anyway. This diff makes that change and we see no measurable performance impact. Reviewed By: jiawenliu64 Differential Revision: D71081537
added 3 commits
March 14, 2025 13:03
Summary: X-link: facebookresearch/FBGEMM#847 Pull Request resolved: pytorch#3766 One of the new interesting changes in the preshuffled F8I4 kernel is that group scales are downcast to FP8. This has the risk of running into dynamic range issues and impacting accuracy. We can mitigate this risk by adding FP32 columnwise scaling to the output. Fortunately, we can do this using EVT so the performance impact is negligible. Differential Revision: D70587477
Summary: X-link: facebookresearch/FBGEMM#855 Pull Request resolved: pytorch#3775 This diff introduces a set of quantization helper functions to fbgemm_gpu/experimental/gen_ai to make it easier to apply the new Int4 packing and preshuffling to weights. Differential Revision: D70643388 Reviewed By: summerdengfb
Summary: Pull Request resolved: pytorch#3800 Working on adding support for stacked mixed dtype grouped gemm with preshuffling. Differential Revision: D70870933
jwfromm
force-pushed
the
export-D71081537
branch
from
1bf1da4 to
84ad32e
Compare
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Mar 14, 2025
Summary: X-link: facebookresearch/FBGEMM#898 Rather than do a separate cumsum operator in torch (which can be expensive), its quite trivial to fuse the sum into the kernel setup we have to do for GEMM anyway. This diff makes that change and we see no measurable performance impact. Reviewed By: jiawenliu64 Differential Revision: D71081537
jwfromm
force-pushed
the
export-D71081537
branch
from
84ad32e to
6505cad
Compare
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Mar 14, 2025
Summary: X-link: facebookresearch/FBGEMM#898 Rather than do a separate cumsum operator in torch (which can be expensive), its quite trivial to fuse the sum into the kernel setup we have to do for GEMM anyway. This diff makes that change and we see no measurable performance impact. Reviewed By: jiawenliu64 Differential Revision: D71081537
|
This pull request was exported from Phabricator. Differential Revision: D71081537 |
jwfromm
force-pushed
the
export-D71081537
branch
from
6505cad to
5715b78
Compare
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Mar 14, 2025
Summary: X-link: facebookresearch/FBGEMM#898 Pull Request resolved: pytorch#3812 Rather than do a separate cumsum operator in torch (which can be expensive), its quite trivial to fuse the sum into the kernel setup we have to do for GEMM anyway. This diff makes that change and we see no measurable performance impact. Reviewed By: jiawenliu64 Differential Revision: D71081537
Summary: X-link: facebookresearch/FBGEMM#898 Pull Request resolved: pytorch#3812 Rather than do a separate cumsum operator in torch (which can be expensive), its quite trivial to fuse the sum into the kernel setup we have to do for GEMM anyway. This diff makes that change and we see no measurable performance impact. Reviewed By: jiawenliu64 Differential Revision: D71081537
|
This pull request was exported from Phabricator. Differential Revision: D71081537 |
jwfromm
force-pushed
the
export-D71081537
branch
from
5715b78 to
11da092
Compare
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: Rather than do a separate cumsum operator in torch (which can be expensive), its quite trivial to fuse the sum into the kernel setup we have to do for GEMM anyway. This diff makes that change and we see no measurable performance impact.
Reviewed By: jiawenliu64
Differential Revision: D71081537