Hello author, may I ask if you ever encountered the error of DDP backpropagation when modifying the code on the basis of stark, such as the picture? I also modified on the basis of stark following your code style, and the error was located at loss.backward(), which could not be done with a single card? I'm very sorry to bother you. Looking forward to your reply.
Hello author, may I ask if you ever encountered the error of DDP backpropagation when modifying the code on the basis of stark, such as the picture? I also modified on the basis of stark following your code style, and the error was located at loss.backward(), which could not be done with a single card? I'm very sorry to bother you. Looking forward to your reply.
