Rui Wang

Hi, Thanks for your interests and your question! This is a crude version of subbatching and as we said in readme, it is going to be udpated very soon (a...

Hi all, Within 48 hours we are going to update the code for memory efficiency and within 72 hours we are going to provide an estimate on the ram usage...

Hi all, Sorry for the delay. It turned out to be _slightly_ more delicate than we expected. But now it runs very long sequences with a much more sensitive to...

Hi @laoshaw, Sorry for this, we seemed to made `--subbatch_size` a necessary argument, which means you have to set `--subbatch_size` explicitly. We are working on a quick fix now. Best

Guess just lower it. For now we do not yet have a rule of thumb, but you can always half the subbatch size each time it runs into bottleneck. We...

We did not write the program to take advantage of two graphics cards though, so we are not sure if this is a specially technology from Nvidia and what happened...

We haven't yet had the time to really tested the runtime with different subbatches yet, but will update the readme soon. However, the GPU memory should not be that low....

Isn't the second card completely idle? By the look of it, it seems we are using the only first card, which looks just fine. The code should only be running...

I see. The model should not take too much out of CPU memory though, so I suppose it is expected. It aligns well with our observation in our machine, that...

But if you feel this is too slow, you could increase the subbatch_size a bit to fully utilize the GPU memory as well.