Skip to content

[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

Open
TarzanZhao wants to merge 2 commits into
modelscope:mainfrom
TarzanZhao:perf/minimax-h3-lora-training
Open

[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:mainfrom
TarzanZhao:perf/minimax-h3-lora-training

MiniMax-H3: selective activation checkpointing, attention output kept

262afc0
Select commit
Loading
Failed to load commit list.

Workflow runs completed with no jobs