Closed poncelettheo closed 2 years ago
Hi! It depends on what you want to do. For inference, a 4GB card should be enough. For training... Well, I trained it on a card with a lot of memory, so I didn't bother cutting the documents. There are some documents that are really big which are very demanding on memory, so you can easily OOM on training. You can try limiting the size of documents during training to reduce the memory consumption. For instance, if a document is [n_seqs, 512], I'd suggest randomly taking 2 out of n_seqs sequences during each epoch.
Thank you very much for your answer ! I have asked for a big enough gpu to train the code however I get a weird error
RuntimeError: CUDA error: CUBLAS_STATUS_EXECUTION_FAILED when calling
cublasSgemm( handle, opa, opb, m, n, k, &alpha, a, lda, b, ldb, &beta, c, ldc)`` but it is related to my environment I think (even if I don't know how to solve this problem)...
I suggest you run the same experiment on cpu, you might get a more understandable error message. If you don't get one at all, then you should try starting with a fresh CUDA installation
Hi! Thanks a lot for uploading your model. I am trying to train the model, but I also get an out of memory error. Would it perhaps be possible to elaborate a bit more on the solution you propose here? I could not directly identify the n_sequences in the code -- in what file could I change this to only taking 2 sequences per document? Thank you so much!
Hello, I want to run your code but as I do not have any GPU I have to ask for one from my laboratory... I was therefore wondering if you had any idea of what GPU memory is needed to run your code because it seams that even 32GB is not enough. Thank you for your understanding.