Add torch compliant grouped gemm API for CK FP8 rowwise #4486
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
|
This pull request was exported from Phabricator. Differential Revision: D78119166 |
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify project configuration. |
cthi
added a commit
to cthi/FBGEMM-1
that referenced
this pull request
Jul 15, 2025
Summary: X-link: facebookresearch/FBGEMM#1543 For PyTorch integration we will need to support several additional cases, as well as leverage slightly different API. This is best observed through the torch test cases, e.g. [test_scaled_grouped_gemm_2d_3d](https://www.internalfb.com/code/fbsource/[fbdb0063f1c1ecca30f5eab8b5341643f680ed51]/fbcode/caffe2/test/test_matmul_cuda.py?lines=1793), [test_scaled_grouped_gemm_3d_2d](https://www.internalfb.com/code/fbsource/[fbdb0063f1c1ecca30f5eab8b5341643f680ed51]/fbcode/caffe2/test/test_matmul_cuda.py?lines=1854) - [Natalias grouped gemm API doc](https://docs.google.com/document/d/1985La6wUUVH1AGBkNhaGKUXzx-9ybtbUp567-vYVOM4/edit?tab=t.0#heading=h.g8lzbjnyzzx9) **A summary is we need these cases:** |**Input Type** | Notes | | 2D-3D | same as fbgemm stacked for MoE | | 3D-2D | not sure use-case for this yet | | 2D-2D | I think this is for backward? | | 3D-3D (BMM) | [Could alternatively leverage FBGEMM BMM kernel](https://www.internalfb.com/code/fbsource/fbcode/deeplearning/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise_batched/) | Pytorch API uses offsets instead of sizes, so we update the kernel setting the grouped gemm parameters to take in offsets as well, and support the above cases. - For BMM we could alternatively leverage the AMD FP8 BMM kernel in FBGEMM. But we can get some "free" support by doing this in the grouped kernel. - I've not yet updated the heuristics to account for the new cases properly. This will come after with a re-tune for generic shapes, as opposed to llama specific. Differential Revision: D78119166
|
This pull request was exported from Phabricator. Differential Revision: D78119166 |
cthi
added a commit
to cthi/FBGEMM-1
that referenced
this pull request
Jul 15, 2025
Summary: X-link: facebookresearch/FBGEMM#1543 For PyTorch integration we will need to support several additional cases, as well as leverage slightly different API. This is best observed through the torch test cases, e.g. [test_scaled_grouped_gemm_2d_3d](https://www.internalfb.com/code/fbsource/[fbdb0063f1c1ecca30f5eab8b5341643f680ed51]/fbcode/caffe2/test/test_matmul_cuda.py?lines=1793), [test_scaled_grouped_gemm_3d_2d](https://www.internalfb.com/code/fbsource/[fbdb0063f1c1ecca30f5eab8b5341643f680ed51]/fbcode/caffe2/test/test_matmul_cuda.py?lines=1854) - [Natalias grouped gemm API doc](https://docs.google.com/document/d/1985La6wUUVH1AGBkNhaGKUXzx-9ybtbUp567-vYVOM4/edit?tab=t.0#heading=h.g8lzbjnyzzx9) **A summary is we need these cases:** |**Input Type** | Notes | | 2D-3D | same as fbgemm stacked for MoE | | 3D-2D | not sure use-case for this yet | | 2D-2D | I think this is for backward? | | 3D-3D (BMM) | [Could alternatively leverage FBGEMM BMM kernel](https://www.internalfb.com/code/fbsource/fbcode/deeplearning/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise_batched/) | Pytorch API uses offsets instead of sizes, so we update the kernel setting the grouped gemm parameters to take in offsets as well, and support the above cases. - For BMM we could alternatively leverage the AMD FP8 BMM kernel in FBGEMM. But we can get some "free" support by doing this in the grouped kernel. - I've not yet updated the heuristics to account for the new cases properly. This will come after with a re-tune for generic shapes, as opposed to llama specific. Differential Revision: D78119166
|
This pull request was exported from Phabricator. Differential Revision: D78119166 |
cthi
added a commit
to cthi/FBGEMM-1
that referenced
this pull request
Jul 16, 2025
Summary: X-link: facebookresearch/FBGEMM#1543 For PyTorch integration we will need to support several additional cases, and leverage slightly different API. This is best observed through the torch test cases, e.g. [test_scaled_grouped_gemm_2d_3d](https://www.internalfb.com/code/fbsource/[fbdb0063f1c1ecca30f5eab8b5341643f680ed51]/fbcode/caffe2/test/test_matmul_cuda.py?lines=1793), [test_scaled_grouped_gemm_3d_2d](https://www.internalfb.com/code/fbsource/[fbdb0063f1c1ecca30f5eab8b5341643f680ed51]/fbcode/caffe2/test/test_matmul_cuda.py?lines=1854) - ngimel's [grouped gemm API doc](https://docs.google.com/document/d/1985La6wUUVH1AGBkNhaGKUXzx-9ybtbUp567-vYVOM4/edit?tab=t.0#heading=h.g8lzbjnyzzx9) **A summary is we need these cases:** |**Input Type** | Notes | | 2D-3D | same as fbgemm stacked for MoE | | 3D-2D | not sure use-case for this yet | | 2D-2D | I think this is for backward? | | 3D-3D (BMM) | [Could alternatively leverage FBGEMM BMM kernel](https://www.internalfb.com/code/fbsource/fbcode/deeplearning/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise_batched/) | Pytorch API uses offsets instead of sizes, so we update the kernel setting the grouped gemm parameters to take in offsets as well, and support the above cases. - For BMM we could alternatively leverage the AMD FP8 BMM kernel in FBGEMM. But we can get some "free" support by doing this in the grouped kernel. - I've not yet updated the heuristics to account for the new cases properly. This will come after with a re-tune for generic shapes, as opposed to llama specific. Differential Revision: D78119166
|
This pull request was exported from Phabricator. Differential Revision: D78119166 |
cthi
added a commit
to cthi/FBGEMM-1
that referenced
this pull request
Jul 16, 2025
Summary: X-link: facebookresearch/FBGEMM#1543 For PyTorch integration we will need to support several additional cases, and leverage slightly different API. This is best observed through the torch test cases, e.g. [test_scaled_grouped_gemm_2d_3d](https://www.internalfb.com/code/fbsource/[fbdb0063f1c1ecca30f5eab8b5341643f680ed51]/fbcode/caffe2/test/test_matmul_cuda.py?lines=1793), [test_scaled_grouped_gemm_3d_2d](https://www.internalfb.com/code/fbsource/[fbdb0063f1c1ecca30f5eab8b5341643f680ed51]/fbcode/caffe2/test/test_matmul_cuda.py?lines=1854) - ngimel's [grouped gemm API doc](https://docs.google.com/document/d/1985La6wUUVH1AGBkNhaGKUXzx-9ybtbUp567-vYVOM4/edit?tab=t.0#heading=h.g8lzbjnyzzx9) **A summary is we need these cases:** |**Input Type** | Notes | | 2D-3D | same as fbgemm stacked for MoE | | 3D-2D | not sure use-case for this yet | | 2D-2D | I think this is for backward? | | 3D-3D (BMM) | [Could alternatively leverage FBGEMM BMM kernel](https://www.internalfb.com/code/fbsource/fbcode/deeplearning/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise_batched/) | Pytorch API uses offsets instead of sizes, so we update the kernel setting the grouped gemm parameters to take in offsets as well, and support the above cases. - For BMM we could alternatively leverage the AMD FP8 BMM kernel in FBGEMM. But we can get some "free" support by doing this in the grouped kernel. - I've not yet updated the heuristics to account for the new cases properly. This will come after with a re-tune for generic shapes, as opposed to llama specific. Differential Revision: D78119166
|
This pull request was exported from Phabricator. Differential Revision: D78119166 |
cthi
added a commit
to cthi/FBGEMM-1
that referenced
this pull request
Jul 16, 2025
Summary: X-link: facebookresearch/FBGEMM#1543 Pull Request resolved: pytorch#4486 For PyTorch integration we will need to support several additional cases, and leverage slightly different API. This is best observed through the torch test cases, e.g. [test_scaled_grouped_gemm_2d_3d](https://www.internalfb.com/code/fbsource/[fbdb0063f1c1ecca30f5eab8b5341643f680ed51]/fbcode/caffe2/test/test_matmul_cuda.py?lines=1793), [test_scaled_grouped_gemm_3d_2d](https://www.internalfb.com/code/fbsource/[fbdb0063f1c1ecca30f5eab8b5341643f680ed51]/fbcode/caffe2/test/test_matmul_cuda.py?lines=1854) - ngimel's [grouped gemm API doc](https://docs.google.com/document/d/1985La6wUUVH1AGBkNhaGKUXzx-9ybtbUp567-vYVOM4/edit?tab=t.0#heading=h.g8lzbjnyzzx9) **A summary is we need these cases:** |**Input Type** | Notes | | 2D-3D | same as fbgemm stacked for MoE | | 3D-2D | not sure use-case for this yet | | 2D-2D | I think this is for backward? | | 3D-3D (BMM) | [Could alternatively leverage FBGEMM BMM kernel](https://www.internalfb.com/code/fbsource/fbcode/deeplearning/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise_batched/) | Pytorch API uses offsets instead of sizes, so we update the kernel setting the grouped gemm parameters to take in offsets as well, and support the above cases. - For BMM we could alternatively leverage the AMD FP8 BMM kernel in FBGEMM. But we can get some "free" support by doing this in the grouped kernel. - I've not yet updated the heuristics to account for the new cases properly. This will come after with a re-tune for generic shapes, as opposed to llama specific. Differential Revision: D78119166
cthi
force-pushed
the
export-D78119166
branch
2 times, most recently
from
89cb88e to
a920a74
Compare
cthi
added a commit
to cthi/FBGEMM-1
that referenced
this pull request
Jul 22, 2025
Summary: X-link: facebookresearch/FBGEMM#1543 For PyTorch integration we will need to support several additional cases, and leverage slightly different API. This is best observed through the torch test cases, e.g. [test_scaled_grouped_gemm_2d_3d](https://www.internalfb.com/code/fbsource/[fbdb0063f1c1ecca30f5eab8b5341643f680ed51]/fbcode/caffe2/test/test_matmul_cuda.py?lines=1793), [test_scaled_grouped_gemm_3d_2d](https://www.internalfb.com/code/fbsource/[fbdb0063f1c1ecca30f5eab8b5341643f680ed51]/fbcode/caffe2/test/test_matmul_cuda.py?lines=1854) - ngimel's [grouped gemm API doc](https://docs.google.com/document/d/1985La6wUUVH1AGBkNhaGKUXzx-9ybtbUp567-vYVOM4/edit?tab=t.0#heading=h.g8lzbjnyzzx9) **A summary is we need these cases:** |**Input Type** | Notes | | 2D-3D | same as fbgemm stacked for MoE | | 3D-2D | needed for backwards | | 2D-2D | needed for backwards | | 3D-3D (BMM) | [Could alternatively leverage FBGEMM BMM kernel](https://www.internalfb.com/code/fbsource/fbcode/deeplearning/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise_batched/) | Pytorch API uses offsets instead of sizes, so we update the kernel setting the grouped gemm parameters to take in offsets as well, and support the above cases. - For BMM we could alternatively leverage the AMD FP8 BMM kernel in FBGEMM. But we can get some "free" support by doing this in the grouped kernel. - I've not yet updated the heuristics to account for the new cases properly. This will come after with a re-tune for generic shapes, as opposed to llama specific. Differential Revision: D78119166
|
This pull request was exported from Phabricator. Differential Revision: D78119166 |
Summary: X-link: facebookresearch/FBGEMM#1543 For PyTorch integration we will need to support several additional cases, and leverage slightly different API. This is best observed through the torch test cases, e.g. [test_scaled_grouped_gemm_2d_3d](https://www.internalfb.com/code/fbsource/[fbdb0063f1c1ecca30f5eab8b5341643f680ed51]/fbcode/caffe2/test/test_matmul_cuda.py?lines=1793), [test_scaled_grouped_gemm_3d_2d](https://www.internalfb.com/code/fbsource/[fbdb0063f1c1ecca30f5eab8b5341643f680ed51]/fbcode/caffe2/test/test_matmul_cuda.py?lines=1854) - ngimel's [grouped gemm API doc](https://docs.google.com/document/d/1985La6wUUVH1AGBkNhaGKUXzx-9ybtbUp567-vYVOM4/edit?tab=t.0#heading=h.g8lzbjnyzzx9) **A summary is we need these cases:** |**Input Type** | Notes | | 2D-3D | same as fbgemm stacked for MoE | | 3D-2D | needed for backwards | | 2D-2D | needed for backwards | | 3D-3D (BMM) | [Could alternatively leverage FBGEMM BMM kernel](https://www.internalfb.com/code/fbsource/fbcode/deeplearning/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise_batched/) | Pytorch API uses offsets instead of sizes, so we update the kernel setting the grouped gemm parameters to take in offsets as well, and support the above cases. - For BMM we could alternatively leverage the AMD FP8 BMM kernel in FBGEMM. But we can get some "free" support by doing this in the grouped kernel. - I've not yet updated the heuristics to account for the new cases properly. This will come after with a re-tune for generic shapes, as opposed to llama specific. Differential Revision: D78119166
cthi
added a commit
to cthi/FBGEMM-1
that referenced
this pull request
Jul 22, 2025
Summary: X-link: facebookresearch/FBGEMM#1543 For PyTorch integration we will need to support several additional cases, and leverage slightly different API. This is best observed through the torch test cases, e.g. [test_scaled_grouped_gemm_2d_3d](https://www.internalfb.com/code/fbsource/[fbdb0063f1c1ecca30f5eab8b5341643f680ed51]/fbcode/caffe2/test/test_matmul_cuda.py?lines=1793), [test_scaled_grouped_gemm_3d_2d](https://www.internalfb.com/code/fbsource/[fbdb0063f1c1ecca30f5eab8b5341643f680ed51]/fbcode/caffe2/test/test_matmul_cuda.py?lines=1854) - ngimel's [grouped gemm API doc](https://docs.google.com/document/d/1985La6wUUVH1AGBkNhaGKUXzx-9ybtbUp567-vYVOM4/edit?tab=t.0#heading=h.g8lzbjnyzzx9) **A summary is we need these cases:** |**Input Type** | Notes | | 2D-3D | same as fbgemm stacked for MoE | | 3D-2D | needed for backwards | | 2D-2D | needed for backwards | | 3D-3D (BMM) | [Could alternatively leverage FBGEMM BMM kernel](https://www.internalfb.com/code/fbsource/fbcode/deeplearning/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise_batched/) | Pytorch API uses offsets instead of sizes, so we update the kernel setting the grouped gemm parameters to take in offsets as well, and support the above cases. - For BMM we could alternatively leverage the AMD FP8 BMM kernel in FBGEMM. But we can get some "free" support by doing this in the grouped kernel. - I've not yet updated the heuristics to account for the new cases properly. This will come after with a re-tune for generic shapes, as opposed to llama specific. Differential Revision: D78119166
cthi
force-pushed
the
export-D78119166
branch
2 times, most recently
from
2b51947 to
4758f24
Compare
|
This pull request has been merged in 53cde4a. |
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:
For PyTorch integration we will need to support several additional cases, as well as leverage slightly different API. This is best observed through the torch test cases, e.g. test_scaled_grouped_gemm_2d_3d, test_scaled_grouped_gemm_3d_2d
A summary is we need these cases:
|Input Type | Notes |
| 2D-3D | same as fbgemm stacked for MoE |
| 3D-2D | not sure use-case for this yet |
| 2D-2D | I think this is for backward? |
| 3D-3D (BMM) | Could alternatively leverage FBGEMM BMM kernel |
Pytorch API uses offsets instead of sizes, so we update the kernel setting the grouped gemm parameters to take in offsets as well, and support the above cases.
Differential Revision: D78119166