Siddharth Narayanan
Results
2
comments of
Siddharth Narayanan
Hi, I see a similar scaling gap when running the tutorial example. I ran the tests with CUDA 11.3, Pytorch 1.11, and an A10 GPU. I installed Triton from source....
@haosdent here's how I fixed the same error in my PPO code. I think this solution will work for trl's PPOTrainer too: Wrap [these lines](https://github.com/huggingface/trl/blob/main/trl/trainer/ppo_trainer.py#L798-L804) with this context manager: ```python...