Closed Hazarch closed 2 months ago
On A800 80G, seting train_batch_size = 20, it seems consume 6 mins using fp16.
I would suggest you to pre-compute all the T5 and VAE features in advance for fast training.
On A800 80G, seting train_batch_size = 20, it seems consume 6 mins using fp16.![1123](https://github.com/PixArt-alpha/PixArt-sigma/assets/124146844/35a03088-bfd2-48b2-80a5-b94d51e4032f)