Kaiyu Shi

The ctc loss will return 0 on errors. One is the ctc loss is not compiled with cuda support but got cuda tensor input. Another situation is input sequence is...

In my case I simply change the `c++11` to `c++14` of `extra_compile_args` in `setup.py`.

@SeanNaren can you review this PR, it's essential to compile correctly against `pytorch-1.6`

Confirmed on pytorch 1.6.0 release. I think the cpp extension API has some BC breaking changes. I'm working on a patch on it, so stay tuned.

@jerryzhang-ss @YongjieYang-YY I solved this problem by re-compiling the pytorch binding after upgrading pytorch. Have you ever tried that before?

The actual name width is calculated as in: https://github.com/wookayin/gpustat/blob/f39d9c9a22e73b574be179e73db70d7b4e4778c2/gpustat/core.py#L435 The name of this option should be something like *min-gpuname-width*. Perhaps it's assumed that user always want the whole name. So...

I don't think nvidia-ml library can report GPU utilization per user. Do you find a solution for that?

whats your interval value

Can you test if `watch -n 1 nvidia-smi` and `watch -n gpustat` use the same amount of CPU time.

@wookayin Good point, I'm working on that.