Marlin Mixed Input Kernel Productionization #3008
Closed
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
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify site configuration. |
|
This pull request was exported from Phabricator. Differential Revision: D61408771 |
|
This pull request was exported from Phabricator. Differential Revision: D61408771 |
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Aug 19, 2024
Summary: Pull Request resolved: pytorch#3008 X-link: facebookresearch/FBGEMM#103 This diff does quite a bit of facelifting to our [Marlin](https://github.com/IST-DASLab/marlin) BF16 X I4 kernels. These improvements include: * Upgrading the kernel with the latest improvements from VLLM. This helps quite a bit with stability and fixes issues with group scaling. * Adds template specializations so that the marlin kernel supports both BF16 and FP16 using a single implementation. * Fixes BF16 Dequantization issue. * Exposes a simplified torch custom op `torch.ops.marlin.marlin_gemm` and convenient helpers for quantizing to the marlin format `marlin_quantize`. * Adds these new ops to our quantize benchmarks. * New tests and better directory structure. One downside of this work is that we have diverged a bit from VLLM so it may be harder to stay in sync going forward. However, I think the benefits of the improvements in this diff outweigh potential sync costs. Reviewed By: jianyuh Differential Revision: D61408771
jwfromm
force-pushed
the
export-D61408771
branch
from
d52b777
to
6cb663c
Compare
|
This pull request was exported from Phabricator. Differential Revision: D61408771 |
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Aug 19, 2024
Summary: Pull Request resolved: pytorch#3008 X-link: facebookresearch/FBGEMM#103 This diff does quite a bit of facelifting to our [Marlin](https://github.com/IST-DASLab/marlin) BF16 X I4 kernels. These improvements include: * Upgrading the kernel with the latest improvements from VLLM. This helps quite a bit with stability and fixes issues with group scaling. * Adds template specializations so that the marlin kernel supports both BF16 and FP16 using a single implementation. * Fixes BF16 Dequantization issue. * Exposes a simplified torch custom op `torch.ops.marlin.marlin_gemm` and convenient helpers for quantizing to the marlin format `marlin_quantize`. * Adds these new ops to our quantize benchmarks. * New tests and better directory structure. One downside of this work is that we have diverged a bit from VLLM so it may be harder to stay in sync going forward. However, I think the benefits of the improvements in this diff outweigh potential sync costs. Reviewed By: jianyuh Differential Revision: D61408771
jwfromm
force-pushed
the
export-D61408771
branch
from
6cb663c
to
2ac7716
Compare
Summary: Pull Request resolved: pytorch#3008 X-link: facebookresearch/FBGEMM#103 This diff does quite a bit of facelifting to our [Marlin](https://github.com/IST-DASLab/marlin) BF16 X I4 kernels. These improvements include: * Upgrading the kernel with the latest improvements from VLLM. This helps quite a bit with stability and fixes issues with group scaling. * Adds template specializations so that the marlin kernel supports both BF16 and FP16 using a single implementation. * Fixes BF16 Dequantization issue. * Exposes a simplified torch custom op `torch.ops.marlin.marlin_gemm` and convenient helpers for quantizing to the marlin format `marlin_quantize`. * Adds these new ops to our quantize benchmarks. * New tests and better directory structure. One downside of this work is that we have diverged a bit from VLLM so it may be harder to stay in sync going forward. However, I think the benefits of the improvements in this diff outweigh potential sync costs. Reviewed By: jianyuh, jiawenliu64 Differential Revision: D61408771
|
This pull request was exported from Phabricator. Differential Revision: D61408771 |
jwfromm
force-pushed
the
export-D61408771
branch
from
2ac7716
to
d5db962
Compare
|
This pull request has been merged in 162cc69. |
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 does quite a bit of facelifting to our Marlin BF16 X I4 kernels. These improvements include:
torch.ops.marlin.marlin_gemmand convenient helpers for quantizing to the marlin formatmarlin_quantize.One downside of this work is that we have diverged a bit from VLLM so it may be harder to stay in sync going forward. However, I think the benefits of the improvements in this diff outweigh potential sync costs.
Differential Revision: D61408771