Re: Enabling ROCm on Everything
On 2023-03-21 21:07, M. Zhou wrote:
> OK. Although I think most users will still use anaconda (including myself),
> we can only see the popcon data after the upload.
Ironically enough, my initial experience with conda was what led me to
hope that we have native support for GPUs in Debian. I remember when
Ampere came out, it sure took a while to get proper support in conda
(although admittedly, pytorch upstream also played a role in that).
Another advantage I see to having proper Debian packages is that
environments can be distributed as container images.
>> For the single binary package TODO list, we'll need to figure out how to
>> deal with shared libraries exceeding 2 GiB.
>
> For your reference.
> The pytorch upstream used some tricks which is toggled by the
> BUILD_SPLIT_CUDA cmake variable. But it turns to be deprecated
> as of the latest pytorch 2.0.0 release.
> https://github.com/pytorch/pytorch/commit/c28cdb53ea1f3e377e478fbdfa64b8cffc3828e6
Interesting, thanks for sharing!
> I have not figured out why the shared object will no longer bloat up.
Reply to: