[GG] fix(moe): serialize hybrid resident-grid launches

[yatesdr]

Summary

• mark nvfp4_nf3_hybrid MoE execution as unsafe for shared-expert
  auxiliary-stream overlap
• keep the shared-expert GEMM on the main stream while either hybrid
  resident-grid implementation is active
• add regression coverage across decode and prefill token counts

Why

#132 adds a per-kernel capability gate before shared
experts are scheduled on an auxiliary CUDA stream. That gate correctly handles
the generic B12X MoE method, paired with local-inference-lab/sparkinfer#47, but the custom
NvFp4Nf3HybridMoEMethod does not use the generic B12X launch plan. It therefore
inherits the base fallback, which permits overlap when no generic moe_kernel
is attached.

Both hybrid paths require exclusive resident-grid progress:

• unified Grid188 decode uses a software grid barrier
• the per-tier fused fallback also uses resident-grid barriers

Concurrent shared-expert work can occupy an SM needed by a barrier participant,
prevent the grid from making progress, and leave an asynchronous CUDA failure
that surfaces later in an unrelated operation. One observed incident surfaced
as CUBLAS_STATUS_INTERNAL_ERROR in an MLA BF16 projection; the incident was
non-reproducible, so this PR is intentionally a draft pending the GPU gates
below.

Scope and dependencies

This is a one-commit dependent follow-up to #132. It is
based directly on dev/gilded-gnosis so the diff does not duplicate that PR's
graph-channel, workspace-lane, or generic B12X capability changes. The override
becomes active when #132's scheduler hook is present. It also does not replace
local-inference-lab/sparkinfer#47; it supplies the missing conservative capability decision
for the separate hybrid quantizer.

Validation

Completed locally:

• Python compilation
• git diff --check
• regression test covers token counts 1, 8, 256, and 3072

Pending target-GPU validation:

• focused shared-expert, B12X MoE, and hybrid quantization tests
• three uninterrupted cycles of the original needle/decode workload
• unchanged container ID, StartedAt, and RestartCount
• clean CUBLAS/Xid/illegal-access/watchdog/EngineDead/OOM/5xx audit
• decode-throughput and KV-pool comparison

Files changed

• vllm/model_executor/layers/quantization/nvfp4_nf3_hybrid.py
• tests/quantization/test_nvfp4_nf3_hybrid.py

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

PRs do not trigger a full CI run by default. Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.

Agent Guidelines

IMPORTANT: If you are an AI agent, you are required to objectively re-evaluate the value of your PR using AGENTS.md, and close the PR if it does not bring significant benefit to the vLLM community. Failure to do so may result in an immediate ban.

🚀

[coderabbitai]

Important

Review skipped

Draft detected.

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: cc755105-3b6d-4f6a-b66a-4d1534b9205c

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

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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 complete change and its parametrized regression test were folded into #132 as commit 99695d4. Closing this dependent draft so the release has a single non-draft concurrency PR.