[perf] MiniMax-H3 LoRA training on a B200 node: 3.74 s to 2.47 s per step at 8 GPUs, mostly from compiling the DiT block body - #1680

[perf] MiniMax-H3 LoRA training on a B200 node: 3.74 s to 2.47 s per step at 8 GPUs, mostly from compiling the DiT block body#1680

TarzanZhao wants to merge 2 commits into

modelscope:mainmodelscope/DiffSynth-Studio:mainfrom

TarzanZhao:perf/minimax-h3-lora-trainingTarzanZhao/DiffSynth-Studio:perf/minimax-h3-lora-trainingCopy head branch name to clipboard