Open 0xdevalias opened 1 year ago
AITemplate is for inference. I don't think it would provide any speedup during training. EDIT: nevermind. You can still accelerate the frozen parts of the model.
Originally posted by @matteoserva in https://github.com/ShivamShrirao/diffusers/issues/32#issuecomment-1275067312
It is just for inference so won't be helpful in training. I also tested it, it's good for inference but also takes a really long time to compile.
Originally posted by @ShivamShrirao in https://github.com/ShivamShrirao/diffusers/issues/32#issuecomment-1279625711
It is just for inference so won't be helpful in training. I also tested it, it's good for inference but also takes a really long time to compile.
Just FYI - The compilation time with the latest open-source version has been improved a lot from our first release. In our experiences, it can be 4X faster for the models where computation-intensive ops are mostly GEMM-family ops. We've made similar improvement for Conv ops in our internal version, which will be sync-ed to the open-source repo later. Stay tuned. Thanks.
Currently there are StableDiffusion examples here:
But it would be awesome if there was also examples of using this for Dreambooth:
See also:
And similar end-goal but these aren't based on AITemplate: