Megatron-DeepSpeed

b227590f - Checking we use fused kernels to compute scaled masked softmax on prefix lm (#209)