Repository navigation
fix: preserve document-relative QSA blocks in packed attention - #236
Closed
taking-lying-flat wants to merge 1 commit into
Closed
taking-lying-flat wants to merge 1 commit into
taking-lying-flat wants to merge 1 commit into
Conversation
taking-lying-flat
force-pushed
the
fix/qsa-packed-block-grid
branch
from
October 20, 2026 13:09
5f9f3eb to
0f4b6be
Compare
Assisted-by: OpenAI Codex
taking-lying-flat
force-pushed
the
fix/qsa-packed-block-grid
branch
from
October 20, 2026 13:16
0f4b6be to
9485780
Compare
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Packed QSA selections use document-relative blocks, but the attention bitmap currently groups tokens by their absolute packed position. If a document starts at an offset not divisible by the compression ratio, sparse attention can include keys the indexer excluded. For example, with document lengths
[3, 20]and block size 4, a selection of[3, 4, 5, 6, 11]expands to[3, ..., 11].Preserve the document's block-grid offset when constructing the bitmap and pass it to the existing kernel offset argument. Forward and backward share tile-mask construction and index compaction, including the physical key tiles crossed by shifted blocks. The compressed bitmap shape stays unchanged.
A local regression, kept outside this PR, compares outputs and Q/K/V gradients against a dense mask built directly from the selected indices. It covers aligned and unaligned documents, multiple block sizes, and forward/backward tile boundaries. Removing either tile-list correction makes the regression fail. This PR changes only the production kernel wrapper.
Validation:
PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 TRITON_INTERPRET=1 OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 .venv/bin/python -m pytest tests/test_qsa_sparse_kernels.py -q: 7 passed, 1 skipped. This executes the Triton kernels through the CPU interpreter; the CUDA long-context test is skipped. The extra regression cases are not part of this PR.flake8,isort --check-only --diff, andyapf --diffon the production file, using the repository-pinned versions: passed.git diff --check: passed.CUDA compilation, GPU performance, and full-model training have not been validated.
Duplicate check: the open QSA PRs #217 and #222 address device registration and CP key-length bounds, respectively. This fixes the block coordinates of packed selections.
AI assistance was used to investigate, implement, and test this change.