Load CLIP and keep large Linear weights stored as FP8 buffers in VRAM (no core patches). Falls back to standard FP16 loading if FP8 is unavailable in the…

A ComfyUI custom node that replaces the stock Load CLIP node with one that stores text encoder weights as FP8 buffers on the GPU, nearly halving persistent VRAM consumption for large models while preserving full inference correctness.

comfyui-fp8-load-clip

clip_name

clip_type_name

device

Load CLIP and keep large Linear weights stored as FP8 buffers in VRAM (no core patches). Falls back to standard FP16 loading if FP8 is unavailable in the current PyTorch build.

Name	Type	Description
clip_name	COMBO	0 options:
clip_type_name	COMBO	18 options: stable_diffusion, stable_cascade, sd3, stable_audio, mochi, cogvideox, +12
deviceopt	COMBO	2 options: default, cpu

Load CLIP FP8

Inputs (3)

Outputs (1)