Megatron-DeepSpeed
e7a12e73 - Turns out there's no issue with the way we build prefix lm

Commit

4 years ago

Turns out there's no issue with the way we build prefix lm

References

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

#213 - Checking we use fused kernels to compute scaled masked softmax on prefix lm

Author

thomasw21

thomasw21

Parents

Loading