fix(rl): reduce Megatron peak memory in selective_log_softmax, causal mask, and discarded logits - #322

fix(rl): reduce Megatron peak memory in selective_log_softmax, causal mask, and discarded logits#322

tastelikefeet merged 3 commits into

modelscope:mainmodelscope/twinkle:mainfrom

Wings-Of-Disaster:fix/megatron-eval-pathWings-Of-Disaster/twinkle:fix/megatron-eval-pathCopy head branch name to clipboard