search

NVIDIA / NeMo-Aligner

Scalable toolkit for efficient model alignment
Apache License 2.0
522 stars 58 forks source link

critic speedup #219

Closed gshennvm closed 2 months ago

gshennvm commented 3 months ago

add some much needed cleanup to the critic and reward model inference servers.

rest of the changes are in the changelog

gshennvm commented 2 months ago

thanks for the review! I checked the numerics locally against the previous main and ran a nemo generate only test. They both look good to my eye so I'm merging now.