guillemgt

icon company DeepL icon location Research Scientist @ DeepL

Results 1 comments of guillemgt

[trainer] fix: PPO reward model resource pool and worker creation with reward loop

> without invoking any methods of worker_group itself This doesn't seem to be the case for vllm rollouts, as in [this line](https://github.com/volcengine/verl/blob/0eb50ec4a33cda97e05ed8caab9c7f17a30c05a9/verl/workers/rollout/vllm_rollout/vllm_async_server.py#L300). > Could you provide some crash details or...