Skip to content

fix: align Kimi verify collector with native state layout - #597

Closed
simone-chen wants to merge 1 commit into
simonec/vllm031-kda-buffer-lifetimefrom
simonec/vllm031-kda-verify-qualification
Closed

simone-chen wants to merge 1 commit into
simonec/vllm031-kda-buffer-lifetimefrom
simonec/vllm031-kda-verify-qualification

Conversation

@simone-chen

Copy link
Copy Markdown
Contributor
  • Kimi 0.31 verify previously failed the qualification gate; its legacy fixture also used zero-based checkpoint IDs, an extra conv-history column and contiguous Q/K/V unrelated to convolution output. Use the native dtype/shape calculators and cache binding, reserve null slot 0, and pass actual packed convolution output to recurrence.
  • Admit only the fixed B200/SD/BF16 input+conv/FP32 recurrence/non-RecoverSSM verify profile. Query width S fixes num_spec=S−1 and accepted=S. Preserve the four registered callback identities, 108 shape points, 216 planned conv/recurrence rows and their separate timing callbacks. Generation, preview and GLM behavior remain unchanged; no schema or consumer changes.
  • Stack on fix: release completed KDA prefill buffers #595. Validate with 73 CPU contract tests and Ruff; callback/getter AST checks confirm unchanged generation, prefill, entry point and both verify timer bodies. Tests cover all four local-head counts, all three widths, cache/input strides, positive IDs, preview behavior and explicit failure propagation.
  • Isolated B200 numerical job 2389279 passed all 36 conditional fixtures; only its six auto verify contexts support this repair. The non-spec fallback controls are dormant in the approved auto profile. Fresh collector smoke/full and the bounded native timing gate remain pending; this draft is not publication approval or historical-data reuse approval.
  • Limits: uniform verify and unpadded per-layer cache pages, without full-model padding, RecoverSSM or mixed scheduler claims. Immutable serving citations are included in the collector and README.

Signed-off-by: Simone Chen <simonec@nvidia.com>
@coderabbitai

coderabbitai Bot commented Oct 26, 2026

Copy link
Copy Markdown

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

  • Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Comment @coderabbitai help to get the list of available commands.

@simone-chen

Copy link
Copy Markdown
Contributor Author

/ok to test 457689d

@simone-chen

Copy link
Copy Markdown
Contributor Author

Superseded by consolidated recurrent/MoE compatibility PR #605, stacked on attention #604. The MoE cleanup is preserved once, including the duplicate in #603. Communication remains #591. Original branches and campaign commit references are preserved; closing without merging.

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