Experimentally round-trips model attention K/V tensors through TQ3. Useful for quality and compression experiments; not a persistent KV cache yet. — ComfyUI…

TQ3 KV cache compression for ComfyUI reducing attention KV cache VRAM by ~4.5x using 3-bit Lloyd-Max quantization with Fast Walsh-Hadamard Transform decorrelation.

ComfyUI-TurboQuant

model

enabled

Experimentally round-trips model attention K/V tensors through TQ3. Useful for quality and compression experiments; not a persistent KV cache yet.

Name	Type	Default	Description
model	MODEL		—
enabled	BOOLEAN	true	—

TurboQuant KV Patch

Inputs (2)

Outputs (1)