I 've use two gpus for the parallel training by "CUDA_VISIBLE_DEVICES=0,1 horovodrun -np 2 dp train --mpi-log=workers input.json" However, both GPUs occupy less than 50 percent, the video memory ...

Our ProLLaMA is the first model to our knowledge capable of simultaneously handling multiple PLP tasks, including generating proteins with specified functions based on the user's intent. EPGF is a ...