vLLM-Omni — output-side multimodal serving

Target: operators who serve image / video / audio / any-to-any generation models with the vLLM-Omni fork of vLLM. vllm-omni extends upstream vLLM (same CUDA/ROCm/NPU/XPU runtime, same OpenAI-compat API server) to add non-autoregressive DiT models, multi-stage pipeline execution, diffusion schedulers, CFG plumbing, and real-time streaming audio I/O — things upstream vLLM does not ship.

This skill is a reference, not a tutorial. SKILL.md holds the mental model, quick-answer router, top pitfalls, and operator cheat sheet. The references/ files hold endpoint catalogs, supported-model tables, stage-config grammar, and the diffusion/DiT details. Read only the reference file that matches the question.

The one thing to know before anything else

vllm-omni is not a fork — it layers on top of upstream vLLM, registers OmniModelConfig, and adds one CLI flag: --omni. Adding --omni to vllm serve routes the server through vllm_omni.entrypoints. As of v0.20.0 the old vLLM entrypoint-hijack / patch.py early-import mechanism was removed — the v0.20.0 release notes state "removal of the old vLLM entrypoint hijack, and runtime changes needed for the 0.20.0 integration path (#3232, #3082, #3352, #3393, #2306)". The omni runtime is now rebased onto upstream vLLM v0.20.0 (rebase PR #3232) rather than monkey-patching it. The architectural claim is to decompose any-to-any models into a graph of disaggregated stages (Thinker / Talker / Code2Wav for Qwen3-Omni; AR-encoder / DiT for Qwen-Image) connected via OmniConnector, so each stage scales independently. The paper (arXiv:2602.02204) claims up to 91.4% JCT reduction vs an unspecified baseline — treat as an architectural argument, not a deployment benchmark.

Version alignment is strict: vllm-omni major.minor must match upstream vLLM major.minor. v0.20.0 (2026-05-07) is the current stable, rebased on upstream vLLM v0.20.0 (CUDA 13.0 / PyTorch 2.11). First stable was v0.14.0 (2026-01-31). Latest pre-release is v0.21.0rc1 (2026-05-25). The v0.19.0rc1 FLUX.1-dev regression (#2730) is fixed in v0.20.0 stable (PR #2760) — no version pin needed anymore.

Quick-answer router

Serving a specific endpoint → references/endpoints.md

• /v1/images/generations, /v1/images/edits (DALL·E-shape)
• /v1/videos (async job) + /v1/videos/sync (raw MP4, 1200s timeout)
• /v1/audio/speech, /v1/audio/voices (list + upload), /v1/audio/speech/batch, /v1/audio/speech/stream (WebSocket)
• /v1/realtime (WebSocket PCM16 in/out for Qwen3-Omni)
• /v1/chat/completions with diffusion via extra_body

Picking a model → references/models.md

• Full supported-architecture → HuggingFace-ID table
• Per-model platform matrix (CUDA / ROCm / NPU / XPU / MUSA)
• Known-issue flags per family

Writing / debugging stage configs → references/stage-config.md

• OmniModelConfig + StageConfig YAML grammar
• OmniConnector types (Shared-memory / Mooncake-Store / Mooncake-Transfer-Engine / RDMA / Yuanrong)
• Pipeline edge validation, entry-point requirement
• stage_id, model_stage, worker_type, engine_output_type, async_chunk

DiT-specific questions → references/diffusion.md

• Schedulers (FlowUniPC + model-specific)
• CFG plumbing (dual CFG for Wan2.2, true_cfg_scale for Qwen-Image, cfg_branch_past_key_values)
• Caches: TeaCache / Cache-DiT / latent cache / noise_pred cache
• Quantization: FP8 (Flux #1640), INT8 (Z-Image/Qwen-Image #1470), GGUF (#1755) — all per-component via ComponentQuantizationConfig
• Ulysses / Ring sequence parallel, CFG-parallel merged-batch TP

Qwen3-Omni realtime + Qwen3-TTS → references/realtime-tts.md

• PCM16 mono @ 16 kHz in / 24 kHz out, OpenAI realtime event shape
• async_chunk: false requirement
• Qwen3-TTS CustomVoice / VoiceDesign / Base modes, 12 Hz / 25 Hz tokenizers
• Voice-upload surface (10 MB cap, consent/ref_text/speaker_description required)

The top operator mistakes this skill exists to prevent

• `/v1/realtime` with `async_chunk: true`. The realtime WebSocket rejects at connection if async_chunk is enabled (api_server.py:1208). Use the default stage-config (vllm_omni/model_executor/stage_configs/qwen3_omni_moe.yaml) — not the ...moe_async_chunk.yaml variant — for realtime sessions. The async-chunk config is for higher-throughput non-realtime Qwen3-Omni serving.

• Qwen3-TTS with CUDA graphs on (v0.18 only). Issue #2866: on v0.18 the code2wav stage crashed when enforce_eager: false, so --enforce-eager was mandatory. #2866 is CLOSED (2026-04-29) and v0.20.0 ships TTS CUDA-graph capture + shared memory pools (release notes cite #2690/#2758/#2803), lifting the requirement. On v0.20.0+ keep --trust-remote-code but --enforce-eager is no longer forced — drop it to regain CUDA-graph throughput, and re-test latency.

• Running the v0.19.0rc1 FLUX artifacts. Issue #2730: FLUX.1-dev generated incorrect images in v0.19.0rc1 (T5 text-encoder bug). Fixed in v0.20.0 stable (PR #2760, merged 2026-04-24). The v0.19.0rc1 tag artifacts are still broken, so do not deploy that specific tag — use v0.20.0+ for any FLUX deployment.

• GLM-Image on v0.18 without `transformers>=5.0`. On v0.18 GLM-Image required a manual pip install 'transformers>=5.0' before serving (the default wheel pinned transformers below 5.0 and GLM-Image silently failed to load). v0.20.0 ships Transformers 5.x compatibility fixes from the upstream rebase — verify whether the manual upgrade is still needed on v0.20.0+ before adding it.

• PCM format on `/v1/realtime`. Qwen3-Omni realtime hard-expects 16-bit PCM mono @ 16 kHz input, outputs PCM at 24 kHz. Stereo, 8 kHz, 24-bit, or WAV-with-header inputs produce garbage or silent failures. Use the reference client in examples/online_serving/qwen3_omni/openai_realtime_client.py as a template.

• Default `guidance_scale=0.0` sentinel. OmniDiffusionSamplingParams treats guidance_scale=0.0 as "not provided" — passing 0.5 intending partial CFG gets coerced. To disable CFG, leave the field unset; to enable, pass > 1.0.

• Prefix caching on a stage that emits latents. Any stage with engine_output_type: latent (thinker stages producing hidden states) must set enable_prefix_caching: false in its engine_args. Prefix cache reuses token-level blocks, which makes no sense for latent outputs — leaving it on surfaces as intermittent stale responses.

• `/v1/videos/sync` for long jobs. The sync endpoint has a hardcoded VIDEO_SYNC_TIMEOUT_S (default ~1200s) and returns 504 past that. Long Wan2.2 / HunyuanVideo-1.5 jobs should use POST /v1/videos (async), then poll GET /v1/videos/{id} and fetch /content.

• Orphan processes after a Wan2.2 crash. Issue #2768: killing one Wan2.2 worker leaves sibling stage processes alive. Wrap launches in a process group + pkill -9 sweep on failure, or use systemd's KillMode=control-group.

• Assuming vllm-omni serves text-only models. If the model has no multimodal output, use stock vLLM — vllm-omni adds overhead for features a text-only model won't exercise, and the community skill explicitly recommends against it. The decision rule: output modality is non-text OR the model name ends -Omni/-Image/-TTS/-Video → vllm-omni; otherwise stock vLLM.

Operator cheat sheet

Install

uv venv --python 3.12 --seed
source .venv/bin/activate

# CUDA — pin upstream vLLM to the matching minor:
uv pip install vllm==0.20.0 --torch-backend=auto

# ROCm:
uv pip install vllm==0.20.0+rocm700 \
  --extra-index-url https://wheels.vllm.ai/rocm/0.20.0/rocm700

# Then the omni package (prebuilt wheel OR editable clone):
uv pip install vllm-omni==0.20.0
# OR: git clone https://github.com/vllm-project/vllm-omni && cd vllm-omni && uv pip install -e .

Python 3.12 is required (3.11 is not supported). Docker image: vllm/vllm-omni:0.20.0.

Serving canonical forms

# Text-to-image (default Z-Image-Turbo quickstart):
vllm serve Tongyi-MAI/Z-Image-Turbo --omni --port 8091

# Qwen-Image with tensor parallelism:
vllm serve Qwen/Qwen-Image --omni --tensor-parallel-size 2 --port 8091

# Qwen3-Omni realtime (default stage config, async_chunk OFF):
vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --omni \
  --tensor-parallel-size 2 --gpu-memory-utilization 0.9 --port 8091

# Qwen3-Omni high-throughput non-realtime (async_chunk ON):
vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --omni \
  --stage-configs-path vllm_omni/model_executor/stage_configs/qwen3_omni_moe_async_chunk.yaml

# Qwen3-TTS (trust-remote-code; --enforce-eager only required on v0.18, lifted by TTS CUDA-graph capture in v0.20.0+):
vllm serve Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice --omni \
  --trust-remote-code --task-type CustomVoice

# Wan2.2 T2V with Ulysses sequence parallel:
vllm serve Wan-AI/Wan2.2-T2V-A14B-Diffusers --omni \
  --ulysses-degree 4 --ulysses-mode strict --port 8091

Common extra flags

Key numbers to memorize

Paired skills

• `vllm-input-modalities` — the complement: text embeddings, reranking, STT (Whisper/Voxtral-STT/Qwen3-ASR), OCR (DeepSeek-OCR). Trigger together when the deployment does both input and output non-text modalities.
• `vllm-nvidia-hardware` — for sizing GB300/NVL72/Rubin capacity for diffusion + CFG-parallel + Ulysses footprints.
• `vllm-caching` — OmniConnector borrows Mooncake from upstream vLLM; the caching skill has the connector-config surface.
• `vllm-observability` — vllm-omni inherits upstream /metrics; profiler hooks (OmniTorchProfilerWrapper) add stage_id + rank awareness to trace files.

Source policy

All claims are cited with file:line, release-note PR refs, or issue IDs. Full anchor list + community channels + third-party plugin catalog in references/sources.md. Compiled 2026-04-18 against v0.18.0; last freshened 2026-05-28 (rebased to v0.20.0 stable; refresh again when the next upstream-rebase release ships).