comfyui-multigpu/web/docs/WanVideoModelLoaderMultiGPU.md

WanVideoModelLoaderMultiGPU

WanVideoModelLoaderMultiGPU wraps the base WanVideo model loader so you can pick both the loader device and the downstream compute device when working with large WanVideo checkpoints. The node patches the underlying WanVideo loader so the model materialises on the device chosen via compute_device while still honouring WanVideo's block swap and quantisation options.

Inputs

Required

Parameter	Data Type	Description
model	STRING	Model file from ComfyUI/models/diffusion_models or ComfyUI/models/unet_gguf to load.
base_precision	STRING	Floating-point format for base weights (fp32, bf16, fp16, or fp16_fast).
quantization	STRING	Optional FP8 quantisation preset; disabled keeps original precision.
load_device	STRING	WanVideo loader slot (main_device or offload_device) used during initial weight materialisation.
compute_device	STRING	MultiGPU device id (e.g. cuda:0, cuda:1, cpu) to run inference on.

Optional

Parameter	Data Type	Description
attention_mode	STRING	Select specialised attention kernels (sdpa, flash_attn_2, flash_attn_3, sageattn, sageattn_3, radial_sage_attention).
compile_args	WANCOMPILEARGS	Torch compile configuration passed through to WanVideo.
block_swap_args	BLOCKSWAPARGS	Enables WanVideo block swapping; supply alongside WanVideoBlockSwapMultiGPU.
lora	WANVIDLORA	Optional Wan LoRA bundle to apply during load.
vram_management_args	VRAM_MANAGEMENTARGS	DiffSynth-Studio memory manager arguments for aggressive VRAM reclamation.
extra_model	VACEPATH	Adds auxiliary model weights (e.g. VACE / MTV Crafter).
fantasytalking_model	FANTASYTALKINGMODEL	Preloads FantasyTalking speech model.
multitalk_model	MULTITALKMODEL	Preloads MultiTalk model.
fantasyportrait_model	FANTASYPORTRAITMODEL	Preloads FantasyPortrait model.
rms_norm_function	STRING	Choose RMSNorm implementation (default Wan variant or pytorch).

Outputs

Output Name	Data Type	Description
model	WANVIDEOMODEL	Initialised Wan diffusion model ready for sampling.
compute_device	MULTIGPUDEVICE	Device id chosen for downstream nodes; pass directly into samplers.