vllm-omni/docs/user_guide/examples/online_serving/diffusers_pipeline_adapter.md

Diffusers Backend Adapter

Source https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/diffusers_pipeline_adapter.

vLLM-Omni supports running diffusion models with the diffusers backend, directly serving any 🤗 Diffusers pipeline online without implementing them natively.

Limitations

The diffusers backend is a black-box adapter. Its primary focus is to serve diffusion models online. Currently, the following features are NOT yet supported. Community contributions are welcome to add these features when they can be implemented with clear Diffusers or vLLM-Omni behavior.

• CFG parallel execution
• Sequence parallel execution
• TeaCache / Cache-DiT acceleration
• Step-wise execution (continuous batching)

For these features, it is recommended to use natively supported pipelines instead.

Model Support

Any model loadable via DiffusionPipeline.from_pretrained() should run, including text-to-image, image-to-image, text-to-video, image-to-video, and text-to-audio.

However, as we strive to ensure output similarity between vLLM-Omni's diffuser backend and plain diffusers library, the following models are particularly verified:

• Qwen/Qwen-Image
• Tongyi-MAI/Z-Image-Turbo
• Wan2.2-I2V-A14B-Diffusers

LTX-2.5 is also supported through a source-preview dependency path described below. The adapter behavior is unit tested, but the gated full checkpoint is not part of the regular CI matrix.

If you find that a model not listed above also produces different outputs from running diffusers model directly. Please consider file an issue.

Usage

vllm serve "stable-diffusion-v1-5/stable-diffusion-v1-5" \
    --omni \
    --diffusion-load-format diffusers

Users turn on the diffusers backend primarily through --diffusion-load-format diffusers argument. There are two more optional arguments, --diffusers-load-kwargs and --diffusers-call-kwargs, which are only valid together with --diffusion-load-format diffusers.

After launching the model, users send a request as usual. Refer to other documentation pages on how to request a particular input/output modality, such as examples/online_serving/text_to_image/openai_chat_client.py.

Configuration Reference

--diffusers-load-kwargs

Passed as-is to DiffusionPipeline.from_pretrained().

This is suitable for model-specific configurations not available through the vLLM-Omni interface. For example: --diffusers-load-kwargs '{"use_safetensors": true}'.

When a parameter is available in the vLLM-Omni interface, it will be adapted here. But if that parameter is simultaneously set in both the vLLM-Omni interface and diffusers_load_kwargs, the latter will take precedence.

--diffusers-call-kwargs

Passed to pipeline.__call__().

This is suitable for sampling parameters not available through the vLLM-Omni interface (such as online serving payloads).

When a parameter is available in the vLLM-Omni interface, it will be adapted here. But if that parameter is simultaneously set in both the vLLM-Omni interface and diffusers_call_kwargs, the former will take precedence (because it is set at request time).

Quantization

Use diffusers_load_kwargs for Diffusers-native quantization options. If diffusers_load_kwargs["quantization_config"] is provided as a dictionary, the diffusers backend builds a Diffusers quantization config and lets Diffusers validate it before calling DiffusionPipeline.from_pretrained().

The diffusers backend also provides a small compatibility shortcut for vLLM-Omni quantization configs when diffusers_load_kwargs does not already contain quantization_config. Currently this maps online/dynamic fp8 and online/dynamic int8 to Diffusers/TorchAO dynamic quantization for transformer components such as transformer or transformer_2. This path requires torchao to be installed.

For example, the CLI can request dynamic FP8 through the vLLM-Omni interface:

vllm serve "Qwen/Qwen-Image" \
    --omni \
    --diffusion-load-format diffusers \
    --quantization-config '{"method": "fp8"}'

Other vLLM-Omni quantization methods, such as gguf, modelopt, mxfp4, mxfp8, serialized checkpoints, static FP8 configs, and layer-name skip lists such as ignored_layers, are intentionally not translated by this compatibility shortcut. Use Diffusers-native configuration through diffusers_load_kwargs or a native vLLM-Omni pipeline for those cases.

Attention Backends

The diffusers backend converts vLLM-Omni standard of attention backend setting to diffusers standard.

Specifically for FLASH_ATTN, it will first attempt to use FlashAttention-3 and then FlashAttention-2.

For each attempted version of FLASH_ATTN and SAGE_ATTN, it will first try to load the attention backend from HuggingFace kernels library, then without.

For unsuccessful attention selection or TORCH_SDPA, it will use the PyTorch's default attention backend.

The loaded attention backend and the failed attempts (if any) are logged to console.

Model Specific Settings

The model loading and inference strictly follows the diffusers library, and they may be different from vLLM-Omni's native interface for some specific models. Users are encouraged to double-check the model pipeline's interface in diffusers' official documentation. Some particular examples are below.

LTX-2.5 (source preview)

The repository linked from the LTX-2.5 release, Lightricks/LTX-2.5, contains component weights and does not have the model_index.json required by DiffusionPipeline.from_pretrained(). Use Lightricks' official converted repository, Lightricks/LTX-2.5-Diffusers, with this backend. Both repositories are gated; accept the model terms on Hugging Face and authenticate with a read token before starting the server.

LTX-2.5 support landed after Diffusers 0.39.0 and is not in a stable Diffusers release yet. After installing vLLM-Omni, install the reviewed upstream commit and Gemma 4-capable Transformers version:

python -m pip install --upgrade 'transformers>=5.10.1' \
  'diffusers @ git+https://github.com/huggingface/diffusers.git@7564fb016dabda0c943416190fc92398c50b1b20'
hf auth login

Serve the default distilled text-to-video pipeline as follows. The model is large; CPU offload and VAE tiling reduce GPU memory pressure at the cost of latency and host memory.

vllm serve "Lightricks/LTX-2.5-Diffusers" \
  --omni \
  --port 8091 \
  --diffusion-load-format diffusers \
  --dtype bfloat16 \
  --enable-cpu-offload \
  --vae-use-tiling \
  --default-sampling-params \
    '{"0":{"height":512,"width":768,"num_frames":121,"fps":24,"frame_rate":24.0,"output_type":"np"}}'

Then request an MP4 with synchronized audio:

curl -X POST http://localhost:8091/v1/videos/sync \
  -F 'prompt=A red fox walks through a snowy forest at dawn while birds sing in the distance.' \
  -F 'size=768x512' \
  -F 'num_frames=121' \
  -F 'fps=24' \
  -F 'seed=42' \
  -o ltx-2.5.mp4

The converted repository selects the distilled transformer by default. When a request supplies neither sigmas, timesteps, nor num_inference_steps, the adapter uses the official eight-sigma distilled schedule and disables CFG, spatio-temporal guidance, and modality guidance. An explicit schedule or step count takes precedence; a step count without sigmas uses Diffusers' generic schedule and may reduce quality. The adapter preserves both frames and the pipeline's singular audio output and reads the audio sample rate from the loaded vocoder (48 kHz for this checkpoint).

This path currently covers the default one-stage distilled T2V pipeline. Loading transformer_full, two-stage upsampling, image/video conditioning, and the diffusion video decoder require additional pipeline assembly and are not enabled by this example. Native vLLM-Omni acceleration features listed under Limitations remain unavailable with the Diffusers backend.

Wan Series

The Wan series video generation models takes boundary_ratio and flow_shift during model initialization (ref), not during inference.

Since our OmniDiffusionConfig contains these two values (source), we can directly pass --boundary-ratio and --flow-shift arguments to vllm serve command.

vllm serve "Wan2.2-T2V-A14B-Diffusers" \
    --omni \
    --boundary-ratio 0.875 \
    --flow-shift 3 \
    --diffusion-load-format diffusers

These extra CLI args will be attempted to pass as-is to the OmniDiffusionConfig dataclass and being accessible during model loading time. Special routines inside the pipeline adapter ensures that they are set properly.