Testing and benchmarking machine translation models

[lynxpda]

I updated Ctranslate2 to the latest current version, replaced the two files in the venv virtual environment and fixed the missing line for GeGLU and that’s it, after that the model conversion with RoPE was successful.
Maybe there is a better way, but it was faster.

[NicoLe]

OK, I’ll try it next week then on a test instance and in the mean time I’ll start running BT with RPE.

[NicoLe]

Hello,
In late May, you mentioned:

@lynxpda

I also had to fix even OpenNMT-py itself to be able to train models with up to 500M parameters with large batch sizes (yes, there is an error in the framework when saving the model to disk during training).

Could you please hand me your fix? It should be in the “model_saver.py” file, but I am clueless as to what might go wrong.

And I want to try an experiment using the 20 enc 20 dec 6k ff (w/ gated-gelu of course ) on my latest datasets to see how it improves the performance over the 161M model.