search

openpsi-project / ReaLHF

Super-Efficient RLHF Training of LLMs with Parameter Reallocation
Apache License 2.0
82 stars 4 forks source link

[Patch] Add a `clear_cache_freq` option in the commandline. #63

Closed garrett4wade closed 1 week ago

garrett4wade commented 1 week ago

This option enables us to manually control the frequency of cleaning CUDA cache during training, avoiding the OOM issue.