Open wanglaofei opened 1 week ago
Unfortunately, block swap doesn't seem to work on multi-GPUs. Multi-GPU training would require DeepSpeed or FSDP, but I don't have time to work on that right now.
Does it mean that flux finetuning is only suitable for one gpu right now? Hope the multi-gpu Flux training comes soon! Thanks for your time, it's a great work.
Can the parameter "--blocks_to_swap" use in multi-gpus settings? Without "--blocks_to_swap", how to finetuning the Flux in multi-gpus with 24GB?