Jiahe Jin

icon company Shanghai Jiaotong University icon location Pittsburgh, PA

Results 2 comments of Jiahe Jin

Help understanding GPU memory spikes during GRPO training.

I encountered the same problem. I tried decrease gpu `actor_rollout_ref.rollout.gpu_memory_utilization` and decrease `actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu `, but they were not working..

Local vlm

I have the same question. If I want to use Azure Deployment, does the framework support evaluation for my local models?