FS#69940 - [python-pytorch-cuda] Missing CUDA capability 5.0 support

Attached to Project: Community Packages
Opened by CUI Hao (cuihao) - Thursday, 11 March 2021, 05:07 GMT
Last edited by Sven-Hendrik Haase (Svenstaro) - Thursday, 11 March 2021, 07:22 GMT
Task Type Bug Report
Category Packages
Status Closed
Assigned To Sven-Hendrik Haase (Svenstaro)
Konstantin Gizdov (kgizdov)
Architecture All
Severity Low
Priority Normal
Reported Version
Due in Version Undecided
Due Date Undecided
Percent Complete 100%
Votes 0
Private No

Details

Description:
python-pytorch-cuda doesn't work on CUDA devices with 5.0 capability. With GTX960M, you end up this error:
"RuntimeError: CUDA error: no kernel image is available for execution on the device"

The Arch package removes support for some old architectures in version 1.6.0:
https://github.com/archlinux/svntogit-community/commit/e8e19559bf0159626994f38b30fe9ac52a17e4e1

But PyTorch upstream still supports capability > 3.0 as of version 1.8.0:
https://github.com/pytorch/pytorch/blob/v1.8.0/torch/cuda/__init__.py#L80

Could you please add these CUDA architectures back? GTX960M is not a very old one.

Additional info:
* package version(s): python-pytorch-cuda 1.8.0-1
This task depends upon

Closed by  Sven-Hendrik Haase (Svenstaro)
Thursday, 11 March 2021, 07:22 GMT
Reason for closing:  Won't implement
Comment by Sven-Hendrik Haase (Svenstaro) - Thursday, 11 March 2021, 07:22 GMT
It's an officially deprecated architecture in cuda and we will not support it. Frankly I'm surprised pytorch still supports it. In any case, they'll be forced to drop support soon ish anyhow. I suggest you compile the package yourself with the architectures you need.

Loading...