ComfyUI-TurboQuant
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
Answer-first: ComfyUI-TurboQuant is an experimental ComfyUI custom node that round-trips attention K/V tensors through TQ3 quantization to study KV-cache compression using 3-bit Lloyd-Max coding and Fast Walsh-Hadamard Transform decorrelation.
TQ3 KV cache compression for ComfyUI. Reduces attention KV cache VRAM by ~4.5x using 3-bit Lloyd-Max quantization with Fast Walsh-Hadamard Transform decorrelation.
Generative-engine profile: llms.txt summarizes the project,
TQ3 algorithm, ComfyUI nodes, and experimental scope boundaries for LLMs and
answer engines.
Motivation
LTX-2.3 22B requires 32.4GB VRAM, barely fitting on a V100 32GB. A large portion of that is the KV cache in transformer attention layers. TurboQuant compresses KV tensors from FP16 (16 bits) to TQ3 (3.5 bits effective), freeing ~5x the KV cache memory.
Installation
cd ~/ComfyUI/custom_nodes/
ln -s ~/ComfyUI-TurboQuant .
Nodes
What is ComfyUI-TurboQuant?
ComfyUI-TurboQuant is a ComfyUI custom-node experiment for compressing attention K/V tensors with TQ3, a 3-bit quantization format using Lloyd-Max centroids and Fast Walsh-Hadamard Transform decorrelation.
Is it a persistent production KV cache?
No. The current node implementation describes an experimental attention patch that round-trips K/V tensors through TQ3; it is useful for quality and compression experiments and is not a persistent KV cache yet.
Which ComfyUI nodes are exposed?
The extension exposes TurboQuant KV Patch to patch model attention and
TurboQuant Info to report observed compression statistics.
TurboQuant KV Patch
Patches a model's attention layers to compress K and V tensors through TQ3 quantization.
- Input: MODEL, enabled (bool)
- Output: MODEL (patched)
TurboQuant Info
Shows compression statistics after inference.
- Input: MODEL
- Output: STRING (stats)
How TQ3 Works
Each 128-float block is compressed to 56 bytes:
- L2 normalize the block
- Fast Walsh-Hadamard Transform (decorrelates values)
- Deterministic random sign flips (spreads energy)
- Absmax scale to [-1, +1]
- Lloyd-Max 8-level codebook quantize (3 bits/value)
- Pack 128 indices into 48 bytes + 4B norm + 4B scale
Round-trip cosine similarity: >0.97 on typical attention vectors.
Self-Test
cd ~/ComfyUI-TurboQuant
python -m tq3_core