FS#60503 - [python-pytorch-cuda] is missing support for compute capability 3.0, 3.2, 3.5
Attached to Project:
Community Packages
Opened by Benoit Brummer (trougnouf) - Saturday, 20 October 2018, 13:18 GMT
Last edited by Sven-Hendrik Haase (Svenstaro) - Monday, 10 December 2018, 16:49 GMT
Opened by Benoit Brummer (trougnouf) - Saturday, 20 October 2018, 13:18 GMT
Last edited by Sven-Hendrik Haase (Svenstaro) - Monday, 10 December 2018, 16:49 GMT
|
Details
pytorch-pytorch-cuda is missing TORCH_CUDA_ARCH_LIST 3.0,
3.2, and 3.5 since 1.0rc1-2. I believe these architectures
are still supported by NVidia.
|
This task depends upon
Closed by Sven-Hendrik Haase (Svenstaro)
Monday, 10 December 2018, 16:49 GMT
Reason for closing: Fixed
Monday, 10 December 2018, 16:49 GMT
Reason for closing: Fixed
Comment by
Sven-Hendrik Haase (Svenstaro) -
Friday, 26 October 2018, 15:17 GMT
Comment by
Benoit Brummer (trougnouf) -
Wednesday, 05 December 2018, 10:49 GMT
Comment by
Sven-Hendrik Haase (Svenstaro) -
Thursday, 06 December 2018, 08:44 GMT
Comment by
Benoit Brummer (trougnouf) -
Thursday, 06 December 2018, 11:03 GMT
Comment by
Sven-Hendrik Haase (Svenstaro) -
Friday, 07 December 2018, 11:11 GMT
Well but it doesn't build with those Archs. Patches appreciated.
Sorry I never took the time to try building for more
architectures. The previous version built fine with
TORCH_CUDA_ARCH_LIST=3.0 but failed when I activated all of them
(I don't know if it has to do with activating multiple
architectures or just 3.2 and/or 3.5 being broken). The current
version doesn't seem to compile anymore which is what brought me
back here (I didn't pay attention to the package updating to
1.0rc1-5 and removing my 3.0 instal), though the git aur package
builds just fine with 3.0.
Well, I'd rather enable newer architecture than older ones. Any
opinions?
1.0rc1-6 compiles fine with export
TORCH_CUDA_ARCH_LIST="3.0;3.2;3.5;3.7;5.0;5.2;5.3;6.0;6.1;6.2;7.0;7.2;7.5"
:)
Can't reproduce using extra-x86_64-build. It fails for me with
linking errors.