[II] Gather sharded Kimi projections through B12X

[voipmonitor]

Behavior

• Routes a one-row BF16 or FP16 Kimi-K3 tensor-parallel projection gather through the public B12X DCP copy collective when VLLM_USE_B12X_DCP_A2A=1.
• Pads each local projection row to the B12X transport contract and removes every rank's padding after the gather, preserving the exact rank-major payload.
• Uses the tensor-parallel coordinator unless the DCP coordinator has the same ordered rank set.
• Preserves tensor-parallel identity at TP1 and the standard vLLM all-gather for prefill, unsupported dtypes, non-CUDA tensors, and unsupported layouts.

Technical reason

The opt-in Kimi-K3 latent-projection sharding implemented by #345 requires every rank to reconstruct the full projection result. The B12X copy collective provides a CUDA-graph-compatible transport for decode without changing the projection's numerical representation.

Compatibility

• This pull request is stacked on [II] TP-shard Kimi MLA latent projections #345.
• The B12X path requires both Kimi-K3 projection sharding and VLLM_USE_B12X_DCP_A2A=1; other model and projection configurations retain the standard collective.
• B12X TP2, TP4, and TP8 use the existing public transport. B12X TP16 requires [II] Enable B12X DCP collectives at world size 16 #338.
• CUDA graph pool ownership is provided separately by [II] Bind B12X TP pools to CUDA graph owners #324 and [II] Bind lazy B12X DCP pools to graph channels #337; composed serving qualification must include those dependencies.

Validation

Executed in voipmonitor/vllm:kimi-k3-qsrt-ii-vllm735952b-b12x180ccab-cu133-torch213-20260815-r3 with the changed source files mounted over the installed tree:

pytest -q tests/models/kimi_k3/test_tp_projection.py tests/models/kimi_k3/test_sequence_parallel.py
35 passed

Ruff formatting, Ruff lint, and git diff --check also pass for all changed files.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

🗂️ Base branches to auto review (1)

• dev/*

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 75b98ecd-94eb-464a-920e-1dfff190552e

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[voipmonitor]

The behavior implemented by this pull request is preserved unchanged in vLLM #385 as commit d872d25970fe. Stable Git patch IDs match. Review and merge #385; this pull request is closed to avoid duplicate review.