release : fix windows hip release

[slaren]

Fixes #13698

[slaren]

@no1wudi @IMbackK I would like to simplify the windows HIP releases, do you know if it would be possible to include all the GPU targets in a single binary, or is it strictly necessary to have a different release for each target? Currently we have different releases for gfx1100, gfx1101, gfx1030.

[IMbackK]

Its not necessary, no. you can have fat binaries.

Im not sure how useful the windows builds are atm really since the list of targets is so small and rather random.
Unfortunately atm rocm on windows is not really supportable as, unlike linux rocm, its very inflexible with the target architectures, to support rdna2 you would have to build for all gfx103x variants for instance, despite these being identical in code generation. This is feasible neither in terms of binary size nor in terms of build time.

The only really feasible way to support windows rocm is to have people self build.

[slaren]

Thanks for the insight. I wonder how other projects handle this. @YellowRoseCx I believe you distribute koboldcpp packages for rocm, can you share some insights of how you handle the different GPU targets in a single release?

[IMbackK]

koboldcpp is built for gfx803, gfx900, gfx906, gfx908, gfx90a, gfx1010, gfx1030, gfx1031, gfx1032, gfx1100, gfx1101 and gfx1102
This list includes various chips you can not actually run on windows anyhow (gfx803, gfx900, gfx906, gfx908, gfx90a) and fails to support various chips you could run on windows like gfx1012, gfx1033, gfx1034, gfx1035, gfx1036, gfx1103, gfx1150, gfx1151, gfx1152, gfx1153, gfx1200 and gfx1201

amd is currenly in the process of introducing generics ie gfx9-generic gfx10.3-generic, gfx11-generic and gfx12-generic that will make this mutch better as then we will only have to build those + gfx1100 and + gfx906, gfx908, gfx90a and gfx942 on linux. They are also in the process of introducing a amdgcn-spirv target to llvm that will allow similar functionality as building the cuda version for ptx

[slaren]

Alright, thanks. I am going to give it a try to include all of these targets in a single binary. The rocblas libraries that are included in our HIP releases are already 900MB, so I figure that even if the size of ggml-hip.dll increases by a factor of 10 it shouldn't change too much the size of the download. If that doesn't work then I guess I will just wait until AMD adds the generic targets.

[slaren]

Adding all of these targets didn't work very well, so instead I added the supported radeon targets listed in the documentation. This results in a 210MB ggml-hip.dll, which is similar to the size of the CUDA backend binary.