componentwise config - Githubissues

chengzeyi / stable-fast

Best inference performance optimization framework for HuggingFace Diffusers on NVIDIA GPUs.

MIT License

1.16k stars 71 forks source link

Open jon-chuang opened 10 months ago

jon-chuang commented 10 months ago

enable vae.encode compile: A100 speedup

original VAE: 1 step LCM sdv1.5 - 24ms speedup! (0.062s -> 0.038s) (I think it doesn't actually work though)
taesd: (4step LCM sdv1.5) - 2ms (0.058s -> 0.056s)

chengzeyi commented 10 months ago