kaldi icon indicating copy to clipboard operation
kaldi copied to clipboard

Pytorchnn support multi-gpus

Open yzmyyff opened this issue 3 years ago • 6 comments

Pytorchnn only uses one gpu. It's very inefficient when training a large amount of text.

yzmyyff avatar Mar 04 '21 10:03 yzmyyff

This is a Kaldi repository. Whet do you mean? K2 is here, btw: https://github.com/k2-fsa/k2.

kkm000 avatar Sep 10 '21 20:09 kkm000

Please see codes in wsj local/pytorchnn. It's another nnlm rescore backend.

yzmyyff avatar Sep 11 '21 00:09 yzmyyff

My best guess is that it's not implemented. A PR is welcome.

@keli78, do you have something to add?

kkm000 avatar Sep 11 '21 02:09 kkm000

@kkm000 Right, we don't have a multi-gpu implementation. I'm afraid I'm not available to add it currently. It will be great if you would like to take it.

keli78 avatar Sep 13 '21 05:09 keli78

@keli78, I'm not up to the task. There is too much work undone on Kaldi that my plate is full, and I have a dayjob (although they do intersect a lot).

@yzmyyff, if you want to develop it, that would be very welcome!

kkm000 avatar Sep 13 '21 10:09 kkm000

I'm sorry that I am not in this direction now.

The conclusion of the previous attempt is that it will obviously reduce server throughput.

So keeping the existing status is enough for me.

yzmyyff avatar Sep 14 '21 02:09 yzmyyff