Decoupling generation and loss batch sizes - #1

Open

sidnarayanan wants to merge 9 commits into

mainfrom

decouple-batch-sizes

Commits on Jan 29, 2025

make sure model is in eval mode before generating
sidnarayanan
committed
9b5a09e
bit more logging
sidnarayanan
committed
bcd74ec
switch to transformers logging
sidnarayanan
committed
8e8b988

Commits on Jan 31, 2025

Merge branch 'main' of https://github.com/huggingface/trl into fix-grpo-generation
sidnarayanan
committed
82d81a5

Commits on Feb 1, 2025

supporting microbatchces for ocmputing loss terms
sidnarayanan
committed
b6a92fe
fix issue with reward func kwargs
sidnarayanan
committed
fd751f3

Commits on Feb 2, 2025

rename arg to micro bsz
sidnarayanan
committed
ed038fe
fix microbatching
sidnarayanan
committed
c763f07
rename to completion_ids in vllm block
sidnarayanan
committed
38c4d63