Skip to content

feat(sdpa): kernels for hy4-flash sparse attention, indexer and engram table reads - #122

Merged
asher merged 10 commits into
mainfrom
feat/sparse-prefill-score-tail
Nov 19, 2026
Merged

asher merged 10 commits into
mainfrom
feat/sparse-prefill-score-tail

Conversation

@asher

@asher asher commented Nov 19, 2026

Copy link
Copy Markdown
Owner

No description provided.

asher added 10 commits November 18, 2026 09:10
…der 256 weights its own thread, 1.7x faster on a 128-wide per-head projection
… on tensor-op gpus, 2.1x to 2.6x faster at a decode step's shapes
…nel, 3.2x faster on a full block, and a lone split writes its output with no partials
…ax kernel, 1.85x faster on a 2048-row chunk of 64 heads
… and superblock pairs evenly over the lanes, 1.15x faster at 1536
@asher
asher merged commit 3b5443c into main Nov 19, 2026
3 checks passed
@asher
asher deleted the feat/sparse-prefill-score-tail branch November 19, 2026 15:47
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant