Skip to content

Harden head-size-256 paged XQA coverage - #32262

Merged
Baiju Meswani (baijumeswani) merged 1 commit into
mainfrom
baijumeswani/xqa-hdim256-test-hardening
Aug 25, 2026
Merged

Baiju Meswani (baijumeswani) merged 1 commit into
mainfrom
baijumeswani/xqa-hdim256-test-hardening

Conversation

@baijumeswani

Copy link
Copy Markdown
Contributor

Addressing open comments from: #32229 (review)

Copilot AI balanced review requested due to automatic review settings August 25, 2026 17:12

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

Hardens CUDA paged XQA head-size-256 tests and addresses prior review gaps.

Changes:

  • Adds FP16/BF16 coverage with INT8/FP8 caches.
  • Strengthens numerical parity and cache-update validation.
  • Skips dispatch assertions when H256 XQA is unavailable.

No actionable issues found.


πŸ’‘ Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

Gate dispatch checks on runtime kernel availability and cover FP16 and BF16 queries with INT8 and FP8 caches.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: a8d618bb-0feb-4d05-a918-93b6accde1b1
@baijumeswani
Baiju Meswani (baijumeswani) force-pushed the baijumeswani/xqa-hdim256-test-hardening branch from 255cb6e to 1396925 Compare August 25, 2026 17:59

@tianleiwu Tianlei Wu (tianleiwu) left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

@baijumeswani
Baiju Meswani (baijumeswani) merged commit d67cc26 into main Aug 25, 2026
91 of 93 checks passed
@baijumeswani
Baiju Meswani (baijumeswani) deleted the baijumeswani/xqa-hdim256-test-hardening branch August 25, 2026 23:12
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.

3 participants