Fixes attention benchmarking and expands test coverage #39
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
Corrects OOM condition logic to require both query and key lengths exceed threshold instead of either one. Captures attention mask output from dynamic mask preparation function and properly passes it to the CUDA kernel instead of using active mask incorrectly. Replaces hardcoded boolean literals with is_causal variable for better code maintainability. Expands benchmark configurations with larger sequence lengths, additional head dimensions, and higher dimensional embeddings to provide more comprehensive performance testing coverage.
Contributor
There was a problem hiding this comment.
Pull Request Overview
This PR refines out-of-memory (OOM) checks, corrects dynamic mask handling in the CUDA kernel call, introduces an is_causal flag for clarity, and expands benchmark configurations for broader performance coverage.
- Tighten OOM logic to require both query and key lengths to exceed the threshold.
- Capture and pass the attention mask (
attn_mask) correctly into the CUDA kernel and addis_causalvariable instead of hardcodedTrue. - Expand benchmark scenarios with larger sequence lengths, additional head dimensions, and higher embedding sizes.
Adds comprehensive memory cleanup with garbage collection and CUDA cache clearing between test configurations to prevent memory issues during extended benchmarking. Expands test configurations to include more diverse scenarios with varying sequence lengths from 4 to 4096 tokens and different head dimensions. Fixes inconsistent attention mask handling by ensuring both Python and CUDA implementations properly use the processed attention mask from prepare_dynamic_mask. Adds proper CUDA synchronization around timing measurements to ensure accurate performance comparisons.
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.
Corrects OOM condition logic to require both query and key lengths exceed threshold instead of either one.
Captures attention mask output from dynamic mask preparation function and properly passes it to the CUDA kernel instead of using active mask incorrectly.
Replaces hardcoded boolean literals with is_causal variable for better code maintainability.
Expands benchmark configurations with larger sequence lengths, additional head dimensions, and higher dimensional embeddings to provide more comprehensive performance testing coverage.