vllm_omni.diffusion.models.diffusers_adapter.pipeline_diffusers_adapter ¶

Diffusers backend adapter for vLLM-Omni.

Provides a black-box wrapper around any 🤗 Diffusers pipeline, enabling vLLM-Omni to directly serve Diffusers models with near-zero per-model code.

The adapter delegates full pipeline execution to diffusers' __call__(). It does NOT support: - CFG parallel (diffusers handles CFG via guidance_scale internally) - Sequence parallel (requires model-specific attention surgery) - TeaCache / Cache-DiT (requires hooking into transformer blocks) - Step-wise execution (continuous batching)

CUDA_FLASH_ATTENTION_BACKEND_ATTEMPTS: list[str] = [
    *CUDA_FLASH_ATTN_3_HUB_ATTEMPTS,
    "_flash_3",
    "_flash_varlen_3",
    *CUDA_FLASH_ATTN_HUB_ATTEMPTS,
    "flash",
    "flash_varlen",
    "_native_flash",
]

CUDA_FLASH_ATTN_3_HUB_ATTEMPTS: list[str] = [
    "_flash_3_hub",
    "_flash_3_varlen_hub",
]

CUDA_FLASH_ATTN_HUB_ATTEMPTS: list[str] = [
    "flash_hub",
    "flash_varlen_hub",
]

logger = logging.getLogger(__name__)

adapter = DiffusersAdapterPipeline(od_config=od_config)
adapter.load_weights()  # calls DiffusionPipeline.from_pretrained()
output = adapter.forward(req)

device = device

od_config = od_config

supports_request_batch = False

supports_step_execution: bool = False

denoise_step(**_: Any) -> Tensor | None

forward(req: DiffusionRequestBatch) -> DiffusionOutput

load_weights() -> None

post_decode(**_: Any) -> Any

prepare_encode(**_: Any) -> Any

step_scheduler(**_: Any) -> None