comfyui-native-quantizer

ComfyUI Native Quantizer is a tool that converts compatible ComfyUI checkpoints and text encoders to ComfyUI's native quantized safetensors formats. It is self-contained, streams large files, preserves unsupported tensors, and validates its output before publishing it. The converter does not modify ComfyUI core and does not produce GGUF, GPTQ, bitsandbytes, AWQ, or other custom-loader formats. It supports a range of native formats including int8, int8-convrot, w4a4, int4-convrot, w4a8, mixed, fp8-e4m3, fp8-e5m2, nvfp4, mxfp8, and precision conversions like fp16 and bf16. Built-in policies cover many ComfyUI diffusion, video, audio, and text-encoder architectures, with experimental support for some models. The tool can be run via command line, with options for inspecting, dry-run, and memory limits. It also supports text encoders with optional quantization of token embeddings. Installation requires the Python environment used by ComfyUI, and the converter accepts local safetensors files, indexed shards, model directories, and trusted local pickle checkpoints. Hugging Face inputs are pinned to a resolved revision and only the requested component is downloaded.