Skip to content

fix: adapt vLLM 0.31 collector runtime contracts - #588

Closed
simone-chen wants to merge 8 commits into
simonec/dynamo-1.6-refreshfrom
simonec/vllm031-scale-qualification
Closed

simone-chen wants to merge 8 commits into
simonec/dynamo-1.6-refreshfrom
simonec/vllm031-scale-qualification

Conversation

@simone-chen

@simone-chen simone-chen commented Oct 26, 2026 •

Copy link
Copy Markdown
Contributor
  • Adapt vLLM 0.31 compute-scale, GDN, KDA and MSA collectors to their native runtime contracts while preserving timing boundaries and paired outputs.
  • Preserve native KDA selection, buffers and recurrence precision. Admit the audited Kimi context/nonspeculative generation profile on B200 and reserve native prefill null cache slot 0. Other layouts and speculative verify remain gated.
  • Adapt MiniMax-M3 MSA metadata/cache bindings and preserve native attention device/host scales during initialization. Keep projection calibration and native decode dispatch.
  • Match MiniMax-M3 MoE serving on 0.31: packed SwiGLU-OAI with checkpoint clamp/alpha/beta, and native ModelOpt mixed-precision resolution through an actual NVFP4 expert prefix. Keep older versions, GPT-OSS layout, case arguments, 27-token sweep and output schema unchanged. Previously mismatched MiniMax data requires fresh qualification.
  • Keep mHC publication gated on fused-operation accounting and architecture identity. No default-version or consumer changes.
  • Validation: 91 focused MoE/GLM/getter CPU tests pass, including native arguments, checkpoint prefix resolution, older-version behavior, and unchanged 3,135-callback/84,645-row inventory. Lint passes. Prior KDA/MSA CPU suites and bounded B200 numerical gates are recorded in the campaign evidence; the new MiniMax MoE binding still requires numerical and native smoke qualification. Collection and reuse approval remain separate gates.

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 5da4619

Signed-off-by: Simone Chen <simonec@nvidia.com>
@simone-chen simone-chen changed the title fix: admit vLLM 0.31 quantization collection fix: prepare vLLM 0.31 scale and recurrent qualification Oct 26, 2026
@simone-chen

Copy link
Copy Markdown
Contributor Author

/ok to test 53fb6dd

Signed-off-by: Simone Chen <simonec@nvidia.com>
@simone-chen simone-chen changed the title fix: prepare vLLM 0.31 scale and recurrent qualification fix: adapt vLLM 0.31 scale and recurrent collectors Oct 26, 2026
@simone-chen

Copy link
Copy Markdown
Contributor Author

/ok to test c8775d0

@simone-chen simone-chen changed the title fix: adapt vLLM 0.31 scale and recurrent collectors fix: adapt vLLM 0.31 collector runtime contracts Oct 26, 2026
Signed-off-by: Simone Chen <simonec@nvidia.com>
@simone-chen

Copy link
Copy Markdown
Contributor Author

/ok to test 667c956

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

Copy link
Copy Markdown
Contributor Author

/ok to test 749bda3

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

Copy link
Copy Markdown
Contributor Author

/ok to test ff1485b

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

Copy link
Copy Markdown
Contributor Author

/ok to test 8cbc7c0

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

Copy link
Copy Markdown
Contributor Author

/ok to test 63d157c

@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