DeepSpeed

bfe7f0db - Fix cudaErrorInvalidConfiguration in attn_softmax() for large seq_length*heads values (#1239)

Fix cudaErrorInvalidConfiguration in attn_softmax() for large seq_length*heads values (#1239) Co-authored-by: Ivan Komarov <dfyz@yandex-team.ru> Co-authored-by: Reza Yazdani <44502768+RezaYazdaniAminabadi@users.noreply.github.com>