vllm-omni
Use vLLM-Omni for omni-modality model inference and serving: local Omni APIs, OpenAI-compatible --omni servers, stage deploy configs, model recipes, and model integration workflows.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- vllm-omni
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Use vLLM-Omni for omni-modality model inference and serving: local Omni APIs, OpenAI-compatible --omni servers, stage deploy configs, model recipes, and model integration workflows.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/install-and-backends.md4.8 KB
- 参考文档references/repo-provenance.md2.3 KB
- 参考文档references/repo-routing-metadata.json505 B
- 参考文档references/troubleshooting.md4.1 KB
- 脚本scripts/check_environment.py6.6 KB
03 SKILL.md 正文
vLLM-Omni
Use this repo skill when a task involves vLLM-Omni: omni-modality inference or serving for text, image, audio, video, action, diffusion, TTS, OpenAI-compatible APIs, stage-based deployment, supported model recipes, or source-level model integration.
Quick start route
-
For install/import/backend checks, read install-and-backends.md and run the safe checker when needed:
python scripts/check_environment.py --require-vllm 0.26Add
--require-cudaonly when live GPU generation or serving must be verified. -
Choose the closest sub-skill:
Task Read Local Python scripts using Omni,AsyncOmni, prompt dictionaries, sampling params, or multimodal outputsoffline-inference vllm serve ... --omni, OpenAI-compatible HTTP payloads, curl/OpenAI SDK clients, realtime/streaming, stage head/headless launchonline-serving Deploy YAML overlays, stages, connectors, memory placement,--stage-overrides, multi-node/stage planningstage-configuration Choosing supported model families, endpoints, hardware backends, diffusion controls, quantization/offload/cache, benchmark metrics model-recipes Adding custom pipelines, model registrations, TTS adapters, or selecting focused maintainer tests model-integration -
If a failure crosses workflows, read troubleshooting.md before running full model examples.
-
To check whether this skill matches the source checkout, read repo-provenance.md.
What vLLM-Omni adds
vLLM-Omni extends vLLM's text-oriented autoregressive runtime with stage-based and diffusion/generation paths for omni-modality workloads:
- multimodal chat and reasoning with text, image, audio, video, and action inputs;
- heterogeneous outputs including text tokens, images, audio, video, latents, trajectories, and custom payloads;
- OpenAI-compatible serving with Omni-specific request fields and endpoints;
- stage-based deploy configs, connectors, head/headless worker launches, and async chunking;
- diffusion image/video/audio pipelines with attention, parallelism, offload, cache, quantization, and LoRA controls;
- TTS and realtime/duplex audio routes;
- model integration surfaces for custom pipelines and supported model families.
Install/backend guardrails
- Use a fresh environment. vLLM, PyTorch, CUDA/ROCm/vendor libraries, media packages, and attention kernels are easy to mix incorrectly.
- Align upstream
vllmwith the vLLM-Omni release line. For this checkout's docs,vllm==0.26.xis the expected upstream line. - Set
VLLM_OMNI_TARGET_DEVICEduring source installs when automation must choose CUDA, ROCm, NPU, XPU, MUSA, or CPU requirements deterministically. - CPU parser/config checks do not prove live model serving. Full generation usually requires GPU/accelerator hardware, model weights, license/cache access, and runtime budget.
- Do not run broad native examples or benchmarks by default; use bundled no-network helpers first.
Common entry points
- Package import:
import vllm_omnifor registration side effects andfrom vllm_omni.entrypoints.omni import Omnifor local generation. - Local API:
Omni(model=..., **kwargs).generate(...)andAsyncOmni(model=...).generate(...). - CLI serving:
vllm serve MODEL --omni ...orvllm-omni serve MODEL --omni .... - Deploy config concepts:
DeployConfig,StageDeployConfig,PipelineConfig,stage_id,devices, connectors,gpu_memory_utilization,async_chunk, platform overrides. - Model/recipe concepts: Qwen3-Omni, Qwen2.5-Omni, Qwen-Image/Z-Image/GLM/Hunyuan/Wan/LTX/Cosmos3/MiniMax-H3, Qwen3-TTS/VoxCPM2/MOSS/Higgs/IndexTTS, DreamZero/GR00T/InternVLA.
Bundled helpers
| Helper | Use |
|---|---|
scripts/check_environment.py |
Verify package metadata, imports, optional CUDA, and Omni CLI help without model downloads. |
sub-skills/offline-inference/scripts/build_offline_request.py |
Generate safe local Python request snippets without importing vLLM-Omni or loading a model. |
sub-skills/online-serving/scripts/build_openai_payload.py |
Generate curl/OpenAI SDK payload scaffolds without sending HTTP requests. |
sub-skills/stage-configuration/scripts/validate_deploy_yaml.py |
Sanity-check deploy YAML shape and connector references without model loading. |
sub-skills/stage-configuration/scripts/plan_stage_memory.py |
Estimate stage placement and memory utilization from user-provided GPU/stage counts. |
sub-skills/model-recipes/scripts/query_model_catalog.py |
Query the bundled representative model catalog by task/backend. |
sub-skills/model-integration/scripts/check_tts_adapter_contract.py |
Statically review TTS adapter source shape without imports or model execution. |
When to refresh this skill
Refresh when vLLM-Omni changes public CLI flags, OpenAI protocol fields, deploy schemas, supported model tables, diffusion controls, TTS adapters, model registry behavior, or aligned upstream vLLM version.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).