ComfyUI Node
TurboWan Model Loader (Quantized)
Load TurboDiffusion quantized models using official inference code
TurboWan Model Loader (Quantized)
- MODEL
◄model_name▾►
◄attention_typesla►
◄sla_topk0.10►
◄offload_modecomfy_native►
Categoryloaders
Inputs (4)
| Name | Type | Default | Description |
|---|---|---|---|
| model_name | COMBO | 0 options: | |
| attention_typeopt | COMBO | sla | Attention mechanism (original=standard, sla=sparse linear attention, sagesla=requires SpargeAttn package) |
| sla_topkopt | FLOAT | 0.100.01–1 | Top-k ratio for sparse attention |
| offload_modeopt | COMBO | comfy_native | comfy_native uses ComfyUI's native async weight offloading (pinned RAM, 2 streams). layerwise_gpu swaps blocks to GPU just-in-time (ComfyUI-style). cpu_only runs the whole forward on CPU (slow). |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |