agentleFS
vllm-projectvllm-project/vllm-omni/.claude/skills/vllm-omni/SKILL.md

vllm-omni

Skill7.8k starsChanged 3 days agoApache-2.0View on GitHub

Deploy, serve, debug, benchmark, and extend multi-stage omni models in vLLM-Omni (Qwen3-Omni, Qwen2.5-Omni, MiniCPM-o, and similar thinker/talker/code2wav pipelines). Use when writing or tuning deploy YAML under vllm_omni/deploy, launching `vllm serve --omni`, diagnosing stage startup hangs, OOM, connector stalls, or broken streaming audio, measuring TTFT/AUDIO_TTFP/AUDIO_RTF with `vllm bench serve --omni`, or navigating the orchestrator and stage runtime code. Routes new omni model integration to `add-omni-model`, omni performance optimization to `omni-perf-opt`, and diffusion, TTS-only, quantization, testing, recipe, and PR tasks to their dedicated skills.

1 other project uses a version of this.

Adopted by

Steady: 0 new projects took it in the last 4 weeks, none the 4 weeks before.thin data

Week of Sep 20: 0 new projectsWeek of Sep 27: 0 new projectsWeek of Oct 4: 0 new projectsWeek of Oct 11: 0 new projectsWeek of Oct 18: 0 new projectsWeek of Oct 25: 0 new projectsWeek of Nov 1: 0 new projectsWeek of Nov 8: 0 new projects
Sep 2026Oct 2026Nov 2026
New projects taking a version of this file each week, over the 8 weeks our latest run read: only 0 in all, so this is a hint, not a trend.

Most recently taken up by

Who's talking about vllm-project/vllm-omni

The biggest 5 of 51 mentions linking to the repository that our crawl found. Each line is their own headline.

See all 51

Talked about on GitHub and Hacker News (51 times).

On GitHub · 50

45 more on GitHub

The file

We split this skill into the rules it gives an agent

2 rules. About 0 are common, 1 is distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Read only the references the task needs.

    line 43
  2. Very specific:

    Localize to a stage: Use stage ids in logs, --log-stats, and --print-stage.

    line 67

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

vLLM-Omni Multi-Stage Omni Models

Entry skill for multi-stage autoregressive omni models. Handle deployment, serving, debugging, and benchmarking here; hand off to the dedicated skill when the task belongs to one.

Mental model

A request flows through a stage graph. For Qwen3-Omni:

stage 0 thinker (LLM_AR)  ->  stage 1 talker (LLM_AR)  ->  stage 2 code2wav (LLM_GENERATION)
  multimodal in, text out       hidden states -> codec codes     codes -> waveform
  • Pipeline (code, models/<model>/pipeline.py) fixes the topology and the stage hand-off functions (model_executor/stage_input_processors/).
  • Deploy YAML (vllm_omni/deploy/*.yaml) sets runtime: devices, memory, batch limits, CUDA graphs, connectors, async_chunk, session_mode, model_runner (v1 / MRv2).
  • Orchestrator (vllm_omni/engine/orchestrator.py) routes each stage output to the next stage or the client; connectors move the data.
  • async_chunk streams partial outputs between stages; it determines first audio packet latency.

Task routing

TaskRead
Understand code paths, find where something livesreferences/architecture.md
Write or tune a deploy YAML, GPU layout, replicasreferences/stage-config.md
Launch a server, call APIs, offline Omnireferences/serving.md
Hang, OOM, crash, stall, bad or broken audioreferences/debugging.md
Benchmark, compare runs, accuracy evals, profilingreferences/benchmarking.md
Qwen3-Omni or Qwen2.5-Omni specificsreferences/models/qwen3-omni.md
MiniCPM-o specificsreferences/models/minicpm-o.md

Read only the references the task needs. For model-specific work, read the model file together with the task reference.

Hand off to dedicated skills

TaskSkill
New thinker/talker/code2wav omni modeladd-omni-model
Omni latency, throughput, streaming smoothness, or GPU-count optimizationomni-perf-opt
New diffusion model / productionizing oneadd-diffusion-model, production-add-diffusion-model
Diffusion performancediffusion-perf-opt
TTS-only modeladd-tts-model
Quantizationquantization
Writing or running tests, CI wiringvllm-omni-test
Recipes under recipes/add-recipe
Self-check before opening a PRprecheck-pr
Reviewing a PRreview-pr
NPU model runner upgradevllm-omni-npu-model-runner-upgrade

Working rules

  1. Resolve the effective config first. Identify the pipeline, the deploy YAML actually loaded (default or --deploy-config), its base_config chain, the platform section, and CLI overrides before reasoning about behavior.
  2. Localize to a stage. Use stage ids in logs, --log-stats, and --print-stage. Isolate the thinker with "modalities": ["text"].
  3. Change one variable at a time and keep the rest of the command identical.
  4. Check CUDA graphs both ways. Compare suspicious output against --enforce-eager (or the stage's enforce_eager) before blaming model code.
  5. Treat source as authoritative. Docs and design notes can lag the code; cite paths and confirm field names in vllm_omni/config/stage_config.py and the model package.
  6. Back claims with evidence. For any performance or correctness claim, record vLLM and vLLM-Omni commits, hardware, deploy YAML, and exact commands, and pair performance results with an accuracy check when numerics may change.

Reporting template

## Summary
[One or two sentences: outcome and root cause or result]

## Setup
- vllm-omni: <commit>, vllm: <commit>, hardware: <GPUs>
- Deploy: <yaml + overrides>
- Command(s): <serve / bench / test>

## Findings
- [Stage-attributed observation with numbers or log evidence]

## Next steps
- [Concrete action]

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.