When translating a sequence of text files using Huggingface, the model is loaded multiple times

[benjaminking]

If you translate a sequence of text files using translate.py and "--start-seq", the model is loaded into memory multiple times. If you translate 10 files, the model will be loaded 10 times, which is inefficient.

[bhartmoore]

Here's a ClearML experiment that gives an example. (I used "--books" instead of the "--start-seq" command, translating from a PT project.) It's very, very common for us to request multiple books at once. In the console log you can see the checkpoints being loaded five times, once for each book requested (1TH, 2TH, 1TI, 2TI, TIT).