Re: Enabling ROCm on Everything
On Tue, 2023-03-21 at 20:48 +0100, Christian Kastner wrote:
> On 2023-03-21 20:38, M. Zhou wrote:
> > If possible, I suggest we stick to only one single binary package that
> > supports multiple selected architectures. The pytorch-rocm popcon
> > will not likely to be very large to deserve such a special treatment.
> > Neither does pytorch-cuda.
>
> Looking at the hype AI currently gets, I wouldn't necessarily say that.
> I think there is a lot of potential -- not just for consumers, but for
> AI-oriented businesses -- in an OS where things just work out of the box.
OK. Although I think most users will still use anaconda (including myself),
we can only see the popcon data after the upload.
> 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
I have not figured out why the shared object will no longer bloat up.
Reply to: