Skip to content

fix: preserve native MoE activation and routing contracts - #589

Closed
simone-chen wants to merge 4 commits into
mainfrom
simonec/vllm031-moe-situ-qualification
Closed

simone-chen wants to merge 4 commits into
mainfrom
simonec/vllm031-moe-situ-qualification

Conversation

@simone-chen

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

Copy link
Copy Markdown
Contributor
  • Pass native SiTU parameters through the callable MoE factory; retain the existing labeled approximation for older factories.
  • Reject declared W4A16 measurements when the selected runtime quantizes activations. Keep the early CT CUTLASS W4A4 guard and inspect both activation descriptors after weight loading on vLLM 0.31.0. GPT-OSS selecting MXFP8 remains an explicit precision gap.
  • Reject the actual selected TrtLlmBf16ExpertsMonolithic leaf on vLLM 0.31.0 when the model requires input routing weights. B200 Llama4 diagnostics show that this native path instead weights expert outputs. Raise before timing without forcing a backend, changing model flags, or dropping queued cases.
  • Preserve output-weighted models, other selected leaf types, older runtime surfaces, existing case identities, and timing methods. Unknown target-runtime precision descriptors remain errors.
  • Validation: 104 CPU getter/case-plan/import/activation/precision/routing guard tests passed; Ruff, format, and diff checks passed. The new routing guard compares the native class object, including a test against an unrelated class with the same name.
  • Native implementation repair and numerical requalification remain required for the blocked path. No GPU jobs were submitted for this change. Full CI is requested for exact head d67e3ce06fc27c1e7cb6c308c9ee5f2f103fc501; GitHub verifies its GPG signature and DCO sign-off is present.
  • CI fixture repair: the shared getter test fake omitted the newly imported BF16 expert module, causing 23 distinct tests to fail across eight unit shards. Add the two strict fake module paths; the pre-fix failure was reproduced locally and all affected suites now pass. Runtime code and campaign source are unchanged by this test-only repair.

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 e4740a6

Signed-off-by: Simone Chen <simonec@nvidia.com>
@simone-chen simone-chen changed the title fix: preserve native SiTU activation in MoE collection fix: preserve native MoE activation and precision contracts Oct 26, 2026
@simone-chen

Copy link
Copy Markdown
Contributor Author

/ok to test 843e3e2

Signed-off-by: Simone Chen <simonec@nvidia.com>
@simone-chen simone-chen changed the title fix: preserve native MoE activation and precision contracts fix: preserve native MoE activation and routing contracts Oct 26, 2026
@simone-chen

Copy link
Copy Markdown
Contributor Author

/ok to test ed5d501

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

Copy link
Copy Markdown
Contributor Author

/ok to test d67e3ce

@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