Fix transformer smaller training vocab

[helpmefindaname]

this PR fixes the usage of the transformer smaller training vocab and improves documentation:

• fix that transformer smaller training vocab are actually used for training and not just temporarily reduce the vocab size before the training starts.
• enforce using a newer version of the library. There are 2 changes:
  • embedding-parameters that are not trainable, but present in the optimizer, will now be rightfully kept not trainable. Hence, running embeddings.model.embeddings.word_embeddings.requires_grad_(False) before training will work with reduce=True.
  • The config rightfully sets the vocab size and therefore reduced models can be saved and later loaded as such.
• the transformer smaller training vocab is now documented in the tutorials
• the onnx tutorial is now slightly improved to fit the newer structure

[alanakbik]

Thanks for this @helpmefindaname! I tested on our cluster - it works with 'distilbert-base-uncased', but running a FLERT training script with 'xlm-roberta-large' throws a weird CUDA error (RuntimeError: CUDA error: CUBLAS_STATUS_NOT_INITIALIZED when calling cublasCreate(handle)``). Can you check?

[helpmefindaname]

Thanks for this @helpmefindaname! I tested on our cluster - it works with 'distilbert-base-uncased', but running a FLERT training script with 'xlm-roberta-large' throws a weird CUDA error (RuntimeError: CUDA error: CUBLAS_STATUS_NOT_INITIALIZED when calling cublasCreate(handle)``). Can you check?

Thanks for pointing this out, this should be fixed with version 0.2.3

[alanakbik]

@helpmefindaname thanks, it works now!