vllm_omni.diffusion.lora ¶

Modules:

DiffusionLoRAManager ¶

Manager for LoRA adapters in diffusion models.

Reuses vLLM's LoRA infrastructure, adapted for diffusion pipelines. Uses LRU cache management similar to LRUCacheLoRAModelManager.

device = device

dtype = dtype

max_cached_adapters = max_cached_adapters

pipeline = pipeline

add_adapter(lora_request: LoRARequest) -> bool

Add a new adapter to the cache without activating it.

list_adapters() -> list[int]

Return list of registered adapter ids.

pin_adapter(adapter_id: int) -> bool

Mark an adapter as pinned so it will not be evicted.

remove_adapter(adapter_id: int) -> bool

Remove an adapter from the cache.

set_active_adapter(
    lora_request: LoRARequest | None,
    lora_scale: float = 1.0,
) -> None

Set the active LoRA adapter for the pipeline.

Parameters:

Name	Type	Description	Default
`lora_request`	`LoRARequest \| None`	The LoRA request, or None to deactivate all adapters.	required
`lora_scale`	`float`	The external scale for the LoRA adapter.	`1.0`