Training Cost
#16
by zhangyang-0123 - opened
Hi,
Would you mind sharing the training hours to finetune the MMDit 4? I am wondering if it is possible for anyone to distill with an affordable budget on cloud.
We trained the final version of (flux-lite-8b) on 8x H200 GPUs for about 140 hours. Plus ~1-2 extra days to pre-compute the input and target hidden states.
Hi, thanks for the awesome project! Could you share about the resolution you used for the distillation? Are you doing it multiple phases (i.e.: 512->1024) or you only do one size or just randomly mixed?