non_blocking=True still fails, lacking corresponding test case about ffcv HOT 8 CLOSED

Fixed on new_ver!

from ffcv.

Hi @andrewilyas,
ffcv is really promising, looking forward to its future!

I'm testing it in a simple application and it looks to me that there's still some problem with the non_blocking feature:

• If I use ToDevice(0, non_blocking=True), I get some CUDA assert complaining that my labels are outside the range [0, n_classes].
• If I use ToDevice(0, non_blocking=False) everything works fine.
• If I don't use ToDevice() at all but use the standard .cuda() from pytorch, everything works fine.

Do you have any idea of what's going on?
I'm using version 0.0.2.

from ffcv.

Sorry you are experiencing issues Can you try v 0.0.3rc1 ?

Do you have an example that we could use to reproduce the bug ?

from ffcv.

Thanks for the reply @GuillaumeLeclerc,

I'll try version 0.0.3rc1 tomorrow in the morning and I'll see if I can prepare a minimum working example.

from ffcv.

It looks like the issue is not present in version 0.0.3rc1.

from ffcv.

Great to hear!

from ffcv.

@lykius This is now in the 0.0.3 final release no need for the RC one.

from ffcv.

Thanks guys!

from ffcv.