feat: modify group-gemm stage number #497
Merged
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
zhyncs
reviewed
Sep 13, 2024
| @@ -85,7 +85,7 @@ cudaError_t CutlassSegmentGEMMWrapper(CutlassSegmentGEMMHandler* handler, DType* | |||
| cutlass::gemm::GemmShape<16, 8, 16>, // Instruction Shape | |||
| cutlass::epilogue::thread::LinearCombination<DType, 8, float, float>, // Epilogue | |||
| cutlass::gemm::threadblock::GemmBatchedIdentityThreadblockSwizzle, // Swizzling Operator | |||
| 8 // Stages | |||
| 4 // Stages | |||
There was a problem hiding this comment.
Adjust the size based on different shared memory instead of directly changing 8 to 4. cc @yzh119
There was a problem hiding this comment.
The value 4 was chosen based on the configuration in cutlass, which should be compatible with most scenarios. 😄
There was a problem hiding this comment.
The example of the cutlass you provided is the SM80, not the SM90.
yzh119
approved these changes
Sep 13, 2024
There was a problem hiding this comment.
LGTM, @jeejeelee thanks for the PR!
I'll merge this at the moment, some todo items:
- add benchmarks for group gemm
- select different configurations according to input shapes and cuda arch.
- add sm90 cutlass group gemm (I have a wip branch but don't have time to work on it at the moment, it will be great if some people in the community can take it over).
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Outdated suggestions cannot be applied.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.
The current group-gemm configuration raises the following error on NVIDIA 3090 :
Modify the stage of group-gemm to 4, reduce the size of dynamic smem, so that it can be called on GPUs like the 3090.
Additionally, I also did a simple comparison on the A800. Modifying the stage to 4 can still slightly improve the performance of group-gemm.
Refer to: https://github.com/NVIDIA/cutlass/blob/main/test/unit/gemm/device/gemm_grouped_sm80.cu