ComfyUI Node
Load Diffusion Model INT4 (W4A4)
Load and Quantize INT4 models with fast comfy-kitchen GPU execution.
Load Diffusion Model INT4 (W4A4)
- pre_lora
- MODEL
◄unet_name▾►
◄weight_dtype▾►
◄model_type▾►
◄on_the_fly_quantizationfalse►
◄enable_convrottrue►
◄lora_modeNone►
Categoryloaders
Inputs (7)
| Name | Type | Default | Description |
|---|---|---|---|
| unet_name | COMBO | 0 options: | |
| weight_dtype | COMBO | Compute dtype. Default follows the model dtype. | |
| model_type | COMBO | Only used for on the fly quantization, to filter sensitive layers. | |
| on_the_fly_quantization | BOOLEAN | false | Quantize a higher precision model to INT4. If the selected model is already INT4 keep unchecked. |
| enable_convrot | BOOLEAN | true | Enable ConvRot for better quantization. |
| lora_mode | COMBO | None | None bakes LoRA patches with normal rounding which is the default behavior. Stochastic bakes with stochastic rounding. Dynamic applies LoRA at inference time. |
| pre_loraopt | PRE_LORA | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |