[Quantization] Add support for Comfy Quants backend

[DN6]

Add support for the Comfy Quant quantization toolkit

Supported formats: FP8 (E4M3/E5M2), INT8 (W8A8, tensorwise), MXFP8, NVFP4, INT4 (SVDQuant W4A4, AWQ W4A16).

comfy-quants is export-only. The inference needs to run with comfy-kitchen

Proposed approach

Add a ComfyQuantConfig / ComfyQuantizer that uses comfy-kitchen to wrap weights as a QuantizedTensor with the appropriate layout.

from diffusers import FluxTransformer2DModel, ComfyQuantConfig

config = ComfyQuantConfig(compute_dtype=torch.bfloat16)
model = FluxTransformer2DModel.from_single_file(
   "path/to/comfy_quant_checkpoint.safetensors",
   quantization_config=config,
)

Related

• comfy-quants https://github.com/Comfy-Org/comfy-quants
• comfy-kitchen https://github.com/Comfy-Org/comfy-kitchen

[PrakshaaleJain]

Hi @DN6, can I take this work up?

[DN6]

@PrakshaaleJain Sure thing 👍🏽 Assigning to you.