Hello, I experimented a program with Hugging Face APIs, especially using the function notebook_launcher() and the environment variable CUDA_VISIBLE_DEVICES, in order to exploit multiple GPUs in my program.
For more details, please refer to the following webpage, including each program and the corresponding log files and demo videos.
https://github.com/unose/masked_lang_model
In short, I was expecting to use 2 GPUs with CUDA_VISIBLE_DEVICES=0,1 (or os.environ[“CUDA_VISIBLE_DEVICES”] = “0,1”) in my program; however, the program failed to do but used just one of them. The output of nvidia-smi was used while running the program. Can you give me any comments or suggestion?