how to translate with multiple instances of the model using multiple GPUs

[StephennFernandes]

hey @pacman100

the translation code could take days given there are multiple SFT datasets and multiple languages to translate on.

is there a way to accelerate the code by launching multiple instances of the model on the same GPU, and launching the same on multiple GPUs such that it can collectively speed up the translation task ?

[StephennFernandes]

found a partial solution to the problem here: huggingface/datasets#6186
still i believe there is a better way to populate all the vram on the single GPU until gpu is full by creating multiple instances of the same model on the same GPU multiple times, and further repeating the same on all available GPUs in the server to maximize flops