[CUTLASS] Add GeMM kernels for Blackwell GPUs #18033
Merged
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
Contributor
Author
|
@tvm-bot rerun |
tqchen
approved these changes
Jun 5, 2025
MasterJH5574
force-pushed
the
tvm-dev/2025-06-02-cutlass-blackwell
branch
2 times, most recently
from
af75beb to
4dd1743
Compare
This PR introduces CUTLASS gemm kernels, groupwise-scaled gemm kernels and group gemm kernels for Blackwell GPUs. Files are reorganized a bit so that the exposed global functions are now architecture agnostic. Prior to this PR, our global function names for CUTLASS kernels usually end with `"_sm90"`, which brings extra complexity when the frontend compiler decides to dispatch kernels when there are multiple supported architectures, such as Hopper and Blackwell. Therefore, this PR renames those global function so that the function names are arch agnostic. During the build time, only the kernels that the specific architecture supports will be built.
MasterJH5574
force-pushed
the
tvm-dev/2025-06-02-cutlass-blackwell
branch
from
4dd1743 to
5f8598e
Compare
MasterJH5574
added a commit
to MasterJH5574/tvm
that referenced
this pull request
Jun 16, 2025
The cutlass kernel build on Hopper GPU was broken since apache#18033. This PR fixes the issue.
tqchen
pushed a commit
that referenced
this pull request
Jun 17, 2025
The cutlass kernel build on Hopper GPU was broken since #18033. This PR fixes the issue.
ShiboXing
pushed a commit
to ShiboXing/tvm
that referenced
this pull request
Aug 10, 2025
This PR introduces CUTLASS gemm kernels, groupwise-scaled gemm kernels and group gemm kernels for Blackwell GPUs. Files are reorganized a bit so that the exposed global functions are now architecture agnostic. Prior to this PR, our global function names for CUTLASS kernels usually end with `"_sm90"`, which brings extra complexity when the frontend compiler decides to dispatch kernels when there are multiple supported architectures, such as Hopper and Blackwell. Therefore, this PR renames those global function so that the function names are arch agnostic. During the build time, only the kernels that the specific architecture supports will be built.
ShiboXing
pushed a commit
to ShiboXing/tvm
that referenced
this pull request
Aug 10, 2025
The cutlass kernel build on Hopper GPU was broken since apache#18033. This PR fixes the issue.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
This PR introduces CUTLASS gemm kernels, groupwise-scaled gemm kernels and group gemm kernels for Blackwell GPUs.
Files are reorganized a bit so that the exposed global functions are now architecture agnostic. Prior to this PR, our global function names for CUTLASS kernels usually end with
"_sm90", which brings extra complexity when the frontend compiler decides to dispatch kernels when there are multiple supported architectures, such as Hopper and Blackwell.Therefore, this PR renames those global function so that the function names are arch agnostic. During the build time, only the kernels that the specific architecture supports will be built.