NVFP4 supoort

[soopasteve89-oss]

Feature Idea

black forest labs have released multiple models in NVFP4 format:
https://huggingface.co/black-forest-labs/FLUX.2-dev-NVFP4
https://huggingface.co/black-forest-labs/FLUX.1-dev-NVFP4
https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev-NVFP4
also lightx2v:
https://huggingface.co/lightx2v/Wan-NVFP4

support for this in comfy would be awesome and greatly appreciated.

Existing Solutions

No response

Other

No response

[rwfsmith]

Can you use Nunchaku to load these?

[soopasteve89-oss]

Nope unfortunately nunchaku spits errors with these models. SVDQ FP4 models are somehow different I guess

[comfyanonymous]

#11635

It's supported now at least for the Nvidia nvfp4 checkpoint like the Flux ones you linked. Lightx format is not supported and will most likely never will be.

[jepjoo]

#11635

It's supported now at least for the Nvidia nvfp4 checkpoint like the Flux ones you linked. Lightx format is not supported and will most likely never will be.

Is this Blackwell only or can it be run on older GPU's that don't have FP4 support?

edit: tried Flux.1-dev with a 4090. It works but is waaay slower than the good old fp8_scaled.

[vimal-v-2006]

Its working slower than fp8 or fp8 mixed

[vimal-v-2006]

but quality is higher than fp8

[soopasteve89-oss]

Thank you team Comfy!

[swampmonster202]

I'm able to load the checkpoint with the load diffusion node, but system locks up or I get OOM in the SamplerCustomAdvanced node. Is there a different workflow or other nodes I should use for NVFP4 models? I'm on a 5090 with 128gb ram on latest Comfy portable build.

[vimal-v-2006]

“Start by running it at 512×512 and let it complete one iteration (20 steps). After that, stop the run, switch the resolution to 1024, and it should work.”

[dur-randir]

Running flux2 on Blackwell nvfp4 is ~28% faster per iteration than q4 gguf.

[swampmonster202]

I got it to run but all that comes out is static. I've got a 5090 with 128gb RAM, running Comfy portable version: 0.8.0, pytorch version 2.8.0+cu129

[kalle07]

16GB VRAM GPU overflow

but qwen fp8 20GB working well

[VyacheslavTeplyakov]

Not working on last desktop version