Skip to content

[WIP][Spec] DSpark compact-verify on DSA backends + sync-free verify-all - #31422

Draft
hnyls2002 wants to merge 11 commits into
mainfrom
lsyin/dspark-compact-dsa-syncfree
Draft

hnyls2002 wants to merge 11 commits into
mainfrom
lsyin/dspark-compact-dsa-syncfree

Conversation

@hnyls2002

@hnyls2002 hnyls2002 commented Jul 16, 2026

Copy link
Copy Markdown
Collaborator

WIP. Enables DSpark compact ragged-verify on GLM-5.2 (DSA) and removes a per-step stream sync on the verify-all path.

  • Capture side now respects supports_ragged_verify_graph: backends without ragged-verify graph support (e.g. DSA) capture fixed-width verify graphs instead of crashing in deepgemm; the epilogue is recorded into those graphs too.
  • Verify-all compact caches the uniform layout and stages verify_lens H2D through pinned memory, dropping the per-step host<->device round-trip.
  • Register DSparkDraftModel so SpecForge dense-draft checkpoints load.

Not final: the pinned-H2D path is only exercised under verify-all so far (measured-SPS-table trimming not yet validated).


CI States

Latest PR Test (Base): ❌ Run #29564493992
Latest PR Test (Extra): ❌ Run #29564493850
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.

@gemini-code-assist

Copy link
Copy Markdown
Contributor

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

@hnyls2002
hnyls2002 force-pushed the lsyin/dspark-compact-dsa-syncfree branch from f5f1efc to 6b01e47 Compare July 16, 2026 23:00
@b8zhong b8zhong mentioned this pull request Aug 5, 2026
41 tasks
@Jiminator Jiminator closed this Sep 14, 2026
@Jiminator
Jiminator deleted the lsyin/dspark-compact-dsa-syncfree branch September 14, 2026 04:42
@alexnails
alexnails restored the lsyin/dspark-compact-dsa-syncfree branch September 14, 2026 05:43
@hnyls2002 hnyls2002 reopened this Sep 14, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants