How about FastHunyuan in GGUF?
No one has done it yet, but many people are waiting for it. Is it possible?
you could try ggc t if you got the gguf-connector from pip and should be able to convert the safetensors to gguf
What is ggc t? Googling didn't help, nothing was found in ComfyUI-Manager for ggc either. I understand there is some node that can save loaded safetensors in GGUF? Or what?
which is a command from gguf-connector
@PixelPlayer for the latest upgrade; our node has new tool to convert safetensors to gguf; you could simply use it with comfyui right away; the converted gguf file will be saved to the output folder just like the picture and video files
That's great! I downloaded your portable version from here https://github.com/calcuis/gguf/releases (because on my Comfy your node was import failed). Load "hunyuan_video_720_fp8_scaled.safetensors" into GGUF Convertor (Alpha) and get the error "Unknown model architecture!". That's all. Maybe I did something wrong?
should do it with bf16, since fp8 is a scaled file; just like gguf quantization should start with the -f16 or -bf16 and cannot do it with 8_0, 5_0 or 4_0; those quantized files loss some key elements for running the process; btw, import failed, seems you don't have gguf-connector; the gguf tensor reader and writer are all from the connector; should get it from pypi or clone the deployment copy to your python site-package folder; the portable version easier
So your 0.0.5 portable version doesn't have gguf-connector installed? Because, I used it. Right now I loaded a simple 11 gb Flux fp8 model, and got 22 gb f16.gguf as output. Apparently it works, but the source model was also fp8, maybe that's the issue. Just don't have fp16 models to check yet, need to download. Thought it would be ok for 6 gb Q4_K_S :) But only got 22 gb.
portable pack has everything, including the gguf-connector; by luck, since some models have the metadata for processing