菜单
Generate and edit images with vLLM-Omni using models like FLUX, Stable Diffusion 3, Qwen-Image, GLM-Image, BAGEL, and Z-Image. Use when generating images from text, editing images, configuring diffusion parameters, or working with image generation models.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Review PRs on vllm-project/vllm-omni by routing to the right domain skills, checking critical evidence, and focusing comments on blocking issues. Use when reviewing pull requests or local branches, triaging review depth, running detailed or default review, or checking tests, benchmarks, and breaking changes in vllm-omni.
Two report kinds; **default output is always HTML** unless the user explicitly asks for Markdown (.md). **Release** — `scripts/compose_full_report.py` (**测试结论**, Buildkite metrics, **Test Result** = Common stack + optional `--log-dir-h*` nightly-style summaries + H100/CI block, **Issue tracking** = GitHub `ci-failure` + *local test* in:title, Open bugs); use `--format markdown` only when the user wants .md or `patch_report_*.py`. **Nightly** — `scripts/nightly_local_log_report.py` from local `nightly_jobs` (fetch: vllm-omni-nightly-local) plus optional latest Buildkite scheduled nightly when token is set; use `--markdown-report` / `--to-stdout markdown` only when the user asks for Markdown. Use when generating Buildkite release summaries, parsing local nightly_jobs with CI cross-check, or opening https://buildkite.com/vllm/vllm-omni/builds?branch=main for CI documentation.
Use before submitting a PR to vllm-project/vllm-omni — self-check the branch against project conventions, catch dead code, verify accuracy/performance claims, and confirm merge readiness. Use when the user says "pre-check", "self review", "pre-submit check", or "check my PR before I open it."
Generate audio and speech with vLLM-Omni using Qwen3-TTS, Fish Speech S2 Pro, CosyVoice3, MiMo-Audio, and Stable-Audio models. Use when synthesizing speech from text, generating audio effects or music, configuring TTS parameters, cloning voices, adding new TTS models, or working with text-to-speech models.
Contribute to vLLM-Omni by adding new model support, fixing bugs, or improving features. Use when integrating a new model into vllm-omni, setting up a development environment, writing tests, or submitting pull requests to the vllm-omni project.
Scale vLLM-Omni across multiple GPUs and nodes using tensor parallelism, pipeline parallelism, OmniConnector disaggregation, connector backends, and Ray. Use when setting up multi-GPU inference, distributing model execution across machines, deploying disaggregated execution, developing OmniConnector backends, or scaling inference horizontally.
| name | vllm-omni-image-gen |
| description | Generate and edit images with vLLM-Omni using models like FLUX, Stable Diffusion 3, Qwen-Image, GLM-Image, BAGEL, and Z-Image. Use when generating images from text, editing images, configuring diffusion parameters, or working with image generation models. |
vLLM-Omni supports text-to-image generation and image editing through diffusion transformer (DiT) models and multi-stage AR+DiT pipelines. Supported model families include FLUX, Stable Diffusion 3, Qwen-Image, GLM-Image, BAGEL, Z-Image, OmniGen2, and more.
| Model | HF ID | Type | Min VRAM |
|---|---|---|---|
| Z-Image-Turbo | Tongyi-MAI/Z-Image-Turbo | Text-to-image | 8 GB |
| Qwen-Image | Qwen/Qwen-Image | Text-to-image (AR+DiT) | 24 GB |
| Qwen-Image-Edit | Qwen/Qwen-Image-Edit | Image editing | 24 GB |
| GLM-Image | zai-org/GLM-Image | Text-to-image | 24 GB |
| BAGEL-7B-MoT | ByteDance-Seed/BAGEL-7B-MoT | Text-to-image + understanding | 24 GB |
| FLUX.1-dev | black-forest-labs/FLUX.1-dev | Text-to-image | 40 GB |
| FLUX.2-klein | black-forest-labs/FLUX.2-klein-4B | Text-to-image | 16 GB |
| FLUX.2-dev | black-forest-labs/FLUX.2-dev | Text-to-image + cache_dit | 24 GB |
| Dreamid-Omni | bytedance/dreamid-omni | Text-to-image (ByteDance) | 24 GB |
| SD 3.5 Medium | stabilityai/stable-diffusion-3.5-medium | Text-to-image | 12 GB |
| OmniGen2 | OmniGen2/OmniGen2 | Text-to-image | 24 GB |
| HunyuanImage3.0 | tencent/HunyuanImage-3.0 | Text-to-image + editing | 40 GB |
Dreamid-Omni from ByteDance and FLUX.2-dev with cache_dit support are available. FLUX.2-klein supports plain string prompts (no dict wrapper needed).
from vllm_omni.entrypoints.omni import Omni
omni = Omni(model="Tongyi-MAI/Z-Image-Turbo")
outputs = omni.generate("a cup of coffee on a table")
outputs[0].request_output[0].images[0].save("coffee.png")
# Start server
vllm serve Tongyi-MAI/Z-Image-Turbo --omni --port 8091
# Generate via API
curl -s http://localhost:8091/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"messages": [{"role": "user", "content": "a sunset over mountains"}],
"extra_body": {
"height": 1024, "width": 1024,
"num_inference_steps": 50,
"guidance_scale": 4.0,
"seed": 42
}
}' | jq -r '.choices[0].message.content[0].image_url.url' \
| cut -d',' -f2 | base64 -d > sunset.png
| Parameter | Description | Typical Range |
|---|---|---|
height | Output height in pixels | 512-2048 |
width | Output width in pixels | 512-2048 |
num_inference_steps | Denoising steps (more = higher quality, slower) | 20-100 |
guidance_scale | CFG scale (higher = more prompt adherence) | 1.0-15.0 |
seed | Random seed for reproducibility | Any integer |
negative_prompt | What to avoid in the image | Text string |
For models that support image editing (Qwen-Image-Edit, LongCat-Image-Edit, HunyuanImage3.0):
from vllm_omni.entrypoints.omni import Omni
omni = Omni(model="Qwen/Qwen-Image-Edit")
outputs = omni.generate(
prompt="Change the sky to sunset colors",
images=["input.jpg"],
)
outputs[0].request_output[0].images[0].save("edited.png")
omni = Omni(model="Tongyi-MAI/Z-Image-Turbo")
prompts = [
"a red car on a highway",
"a blue ocean with waves",
"a green forest in autumn",
]
outputs = omni.generate(prompts)
for i, out in enumerate(outputs):
out.request_output[0].images[0].save(f"image_{i}.png")
Note: diffusion pipeline max_batch_size defaults to 1. Input lists are processed sequentially unless you modify stage configs to increase batch size.
Black or blank images: Increase num_inference_steps. Some models need 50+ steps for good results.
Out of memory: Reduce resolution, or use CPU offloading:
vllm serve <model> --omni --cpu-offload-gb 10
Slow generation: Enable TeaCache for 1.5-2x speedup (see vllm-omni-perf skill). Multi-thread weight loading (enabled by default for diffusion models) also reduces startup time significantly.
HunyuanImage3.0 garbage output in offline inference: Fixed in #3243. The AR stage now uses the Instruct chat template (User:/Assistant: framing) instead of the pretrain format. Trigger tags (💭, <recaption>) must go after Assistant:, not before the user prompt. Use build_prompt_tokens() from vllm_omni.diffusion.models.hunyuan_image3.prompt_utils for segment-by-segment tokenization that avoids cross-segment BPE merges. MoE routing now runs in fp32 (matching HF). VAE pixel values must stay fp32 through preprocessing — do not pre-cast to bf16.
HunyuanImage3.0 load_weights error: Fixed in #1598. Ensure you are using the latest vllm-omni.
HunyuanImage3 online/offline output mismatch: Fixed in #3500/#3516. Online multistage path now uses build_prompt_tokens() matching offline behavior.
HunyuanImage3 AR sampler fails with batched requests: Fixed in #3590. max_num_seqs > 1 now supported for AR sampling.
HunyuanImage3 deploy config fails at startup: Fixed in #3537. Pipeline name changed to hunyuan_image_3_moe; update any custom deploy YAMLs.
HunyuanImage3 KV reuse broken under sequence parallel: Fixed in #3546. ar_kv_reuse_len is now propagated through the DiT forward pass.
Diffusers backend crash on models without model_index.json: Fixed in #3644. Non-diffusers-format models now emit a warning instead of crashing.
GLM-Image filepath errors: Fixed in #1609. Models with model_subdir or tokenizer_subdir now resolve paths correctly.
BAGEL think mode for text2text/img2text: BAGEL now supports reasoning/thinking mode for text-to-text and image-to-text modalities via --think flag. Injects VLM_THINK_SYSTEM_PROMPT to enable chain-of-thought output. Fixed in #2503.
Qwen-Image tiny request sizes: Small requests (below VAE alignment) are now clamped to minimum valid dimensions instead of collapsing to zero. All Qwen-Image pipeline variants use the shared normalize_min_aligned_size() helper. Fixed in #2637.