Simple quantization scripts for ComfyUI diffusion models

"Quick Start"

1. Activate ComfyUI's venv.

2. Copy config from configs/ and customize yours.
   Rules are matched top-to-bottom (first match = highest priority/fallback)

   {
     "format": "comfy_quant",
     "block_names": ["net.blocks."],
     "rules": [
       { "policy": "keep", "match": ["blocks.0."] },
       { "policy": "float8_e4m3fn", "match": ["v_proj", "adaln_modulation", ".mlp"] },
       { "policy": "nvfp4", "match": ["k_proj", "q_proj", "output_proj"] }
     ]
   }

3. Run quantize.py with:

   python quantize.py your_config.json model.safetensors output.safetensors

---

NVFP4 Calibration (in a clumsy way)

NVFP4 requires input scale calibration.

1. Setup ComfyUI:

   • Apply patches from patch/ directory to ComfyUI
   • Copy patch/extra/calibration.py → ComfyUI/extra/
   • Run ComfyUI with:

     COMFY_CALIB=1 python3 main.py ...

2. Generate calibration data:

   • Use your NVFP4-quantized model for normal image generation
   • This creates CALIB_DATA.json

3. Apply calibration:

   python add_input_scale.py CALIB_DATA.json uncalibrated_model.safetensors output.safetensors
   

4. Verify: Test generation with the calibrated model.

Credits

utils/convrot.py

Group-wise Hadamard rotation for INT8 quantization quality improvement.
Uses the Kronecker construction from ConvRot (2025) Theorem 3.3.

Originally from: https://github.com/newgrit1004/ComfyUI-ZImage-Triton
License: MIT