chengzeyi
Potential improvements to stable-fast
Hi chengzeyi, Wanted to first off congratulate you on this awesome work! I have actually also been working on a similar project here but I have recently stopped development since your project has already been widely adopted. However, there are some features that I have been working on that I believe could enhance stable-fast
- Using torch.fx to overwrite and accelerate unet
- Supporting tensor parallelism for people with more than 1 gpu
- INT4 quantization with GPTQ
- Sparse inference
I would love to work with you on the above topics if possible since I have already partially implemented quite a few of these! Please let me know if you could see us collaborating in the future..
@arnavdantuluri It would be great! I don't have more than one GPU so haven't even considered tensor parallelism. And writing fx passes is more complicated than torchscript so I haven't implemented them yet. I would take a look at your project. It must have a lot of things worth leaning for me!
Awesome! Very excited to work with you @chengzeyi. Do you happen to have a Discord handle I could add? I think that would make communication much easier..
@arnavdantuluri Aha, in fact, as you can see, we have a Discord server here: https://discord.gg/kQFvfzM4SJ