ComfyUI Node

Load CLIP INT8 (W8A8)

Load an INT8 text encoder with fast triton inference.

By DrBearJew·Created 25 days ago·Updated 11 days ago· 0
Load CLIP INT8 (W8A8)
    • CLIP
    clip_name
    type
    weight_dtype
    on_the_fly_quantizationfalse
    enable_convrottrue
    Categoryloaders

    Inputs (5)

    NameTypeDefaultDescription
    clip_nameCOMBO0 options:
    typeCOMBO28 options: stable_diffusion, stable_cascade, sd3, stable_audio, mochi, ltxv, +22
    weight_dtypeCOMBOINT8 compute dtype. 'default' follows the encoder dtype.
    on_the_fly_quantizationBOOLEANfalseQuantize a bf16/fp16 encoder to INT8 at load. Leave OFF for an already-INT8 file.
    enable_convrotBOOLEANtrueConvRot rotation for on-the-fly quant. Pre-quantized files carry their own convrot flag per-layer and ignore this.

    Outputs (1)

    NameTypeDescription
    CLIPCLIP