Support for candidate generation and tuning attention#2743
Open
keshavvinayak01 wants to merge 13 commits intomainfrom
Open
Support for candidate generation and tuning attention#2743keshavvinayak01 wants to merge 13 commits intomainfrom
keshavvinayak01 wants to merge 13 commits intomainfrom
Conversation
Signed-off-by: Keshav Vinayak Jha <keshavvinayakjha@gmail.com>
Signed-off-by: Keshav Vinayak Jha <keshavvinayakjha@gmail.com>
Contributor
Coverage report
This PR does not seem to contain any modification to coverable code. |
Signed-off-by: Keshav Vinayak Jha <keshavvinayakjha@gmail.com>
Signed-off-by: Keshav Vinayak Jha <keshavvinayakjha@gmail.com>
Signed-off-by: Keshav Vinayak Jha <keshavvinayakjha@gmail.com>
Signed-off-by: Keshav Vinayak Jha <keshavvinayakjha@gmail.com>
keshavvinayak01
changed the title
keshavvinayak01
commented
Jan 16, 2026
Author
keshavvinayak01
left a comment
keshavvinayak01
left a comment
There was a problem hiding this comment.
@bangtianliu It's a WIP, but if you could go through the changes and see if they make sense.
All these changes were required to make the boo_tuner work with attention, as well as generate more and better candidates for benchmarking.
Contributor
|
Nice! I like the changes. Will redirect to @bangtianliu for review. |
bangtianliu
requested changes
Jan 16, 2026
Comment on lines
899
to
907
| # For attention ops, use VectorDistribute pipeline instead of TileAndFuse | ||
| if dispatch_tuner.get_dispatch_kind() == common.DispatchKind.attention: | ||
| if args.codegen_pipeline != CodegenPipelines.llvmgpu_vector_distribute: | ||
| logging.info( | ||
| f"Attention operation detected. Overriding codegen pipeline " | ||
| f"from {args.codegen_pipeline} to llvmgpu_vector_distribute" | ||
| ) | ||
| args.codegen_pipeline = CodegenPipelines.llvmgpu_vector_distribute | ||
|
|
Contributor
There was a problem hiding this comment.
This code is unnecessary. Since we already have code to handle this case
amd-shark-ai/amdsharktuner/amdsharktuner/constraint_generator.py
Lines 653 to 658 in cab3ead
Author
There was a problem hiding this comment.
The referenced code acts like a guard instead of returning solutions. We should either change that code to overwrite vector-distribute for the attention case or keep my code.
Contributor
|
Left some comments here. Can do another pass once this PR is ready for review |
keshavvinayak01
changed the title
Extract duplicated bf16/f16 with f32 accumulator compatibility logic from common.py and dispatch_constraints.py into a shared helper function. Use isinstance() for type comparison instead of str(). Revert benchmark log level from info back to debug. Add test for the new helper. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> Signed-off-by: Keshav Vinayak Jha <keshavvinayakjha@gmail.com>
keshavvinayak01
changed the title
Fix denorm_flushing type to list[bool] with default [False] and use local variable for attention pipeline selection instead of mutating args. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> Signed-off-by: Keshav Vinayak Jha <keshavvinayakjha@gmail.com>
Signed-off-by: Keshav Vinayak Jha <keshavvinayakjha@gmail.com>
Signed-off-by: Keshav Vinayak Jha <keshavvinayakjha@gmail.com>
keshavvinayak01
changed the title
…xed accumulator type matching in MMA compatibility check Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> Signed-off-by: Keshav Vinayak Jha <keshavvinayakjha@gmail.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> Signed-off-by: Keshav Vinayak Jha <keshavvinayakjha@gmail.com>
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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.
No description provided.