[simone-chen]

Changes

• Repair vLLM 0.31 attention collectors that fail or use stale layouts when native BMM outputs, attention metadata, or shared KV-cache allocation change. Preserve native selectors, production case grids, and timing boundaries.
• Build MLA/DSA metadata through the serving AttentionGroup factory, including the indexer block-table width. Support the native FlashInfer, FlashAttention, and Triton dense-attention routes.
• Preserve the full model CSA indexer page-layout constraint when collecting isolated FlashMLA HCA, so native MLA cache unification can allocate the correct shared layout.
• Initialize ModelOpt NVFP4 W4A4 module weights through vLLM's unified ModelOptLinearMethod and native quantizer. Match the exact weight/activation keys and ctx.group_size; retain the older method interface without including W4A16 or other formats.
• Include a hash-bound, opt-in campaign runtime manifest and fail-closed version-route parsing. Keep existing defaults and older routes unchanged.

Validation

• The NVFP4 repair passes 18 input-population tests plus 48 related MLA, DSA, cache, graph, and HCA compatibility tests. The HCA allocator repair also passes all 25 focused DSV4 compatibility tests.
• Full CI passed for exact head 1ee0d025d17d5a1b77f120e1bf65ff986671363b; GitHub verifies its GPG signature and DCO sign-off. No new head has been pushed since that run.
• In the composed B200 campaign checkout 9b4de9dbc3fa721af0a3965d16f04d3c5dce89bf, all 14 bounded BF16/tensor-FP8 fused A fixtures pass native eager and CUDA graph checks across DeepSeek-V3, DeepSeek-V3.2, and GLM-5.2. These checks cover the fused projection and FP8 shard-scale contracts only.
• ModelOpt NVFP4, block-FP8 partial widths, kv_b conversion, and remaining model-specific attention paths retain their separate GPU qualification gates. Successful numerical probes do not establish timing reuse, complete collection coverage, or end-to-end prediction accuracy.

Dependencies and scope

• Draft PR stacked on fix: qualify TRT-LLM rc29 collectors for Dynamo 1.6 #568, with the route-parser fix from ci: support version-routed collector change detection #592 included. Final publication targets main; this change does not promote performance-data defaults.

[coderabbitai]

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]

/ok to test 079052e

[simone-chen]

/ok to test bb77a29

[simone-chen]

/ok to test bb77a29

[simone-chen]

/ok to test 389708b

[simone-chen]

/ok to test 42ee6e0

[simone-chen]

/ok to test 5852b28

[simone-chen]

/ok to test 11949f7

[simone-chen]

/ok to test b1c756e

[simone-chen]

/ok to test 13be217

[simone-chen]

/ok to test 3921c9c

[simone-chen]

/ok to test 364330c

[simone-chen]

/ok to test 1ee0d02

[simone-chen]

Superseded by consolidated attention compatibility PR #604 for Tianhao's review. The selected review set is #604, #605, and #591. Original branch and campaign commit references are preserved; closing without merging.