fix: guard softmax_offset.grad.zero_() by is_training - #3389

fix: guard softmax_offset.grad.zero_() by is_training#3389

andrewwhitecdw wants to merge 2 commits into

NVIDIA:mainNVIDIA/TransformerEngine:mainfrom

andrewwhitecdw:bugfix/run-attention-with-cp-guard-softmax-offset-grad-zero-by-isandrewwhitecdw/TransformerEngine:bugfix/run-attention-with-cp-guard-softmax-offset-grad-zero-by-isCopy head branch name to clipboard