I updated easyr1 in my codebase recently and I found two changes I can not understand
- it seems that you guys removed max_num_seqs for vllm.. why???
- it seems that warmup must be used and we only do warmup but no lr decay anymore. I tried to disable warmup but without touching the source code I can't find a way to make 0th step lr in wandb non 0.
Both things are working in old repo, why do this simplification? or it's because you guys need to change something to adapt the new version of packages?
I updated easyr1 in my codebase recently and I found two changes I can not understand
Both things are working in old repo, why do this simplification? or it's because you guys need to change something to adapt the new version of packages?