Open durdin85 opened 2 years ago
I've just tried my tensorflow code and also benchmark with stock AMD rocblas 5.1.1 and it works, so I believe there's no need to install patched rocblas on 5.1.1 any more. Looking for kernels, there are gfx803 files on AMD deb.
If you run python in venv and add the library to path export LD_LIBRARY_PATH=/opt/rocm/lib
to this there might be not need to tamper with system files any more.
@durdin85 seems r9 fury and r9nano didnot have issues. I havenot test them, cannot confirm it.
I think 2.43.999 is a good idea, I will have a try.
@durdin85 Add patch version to deb package version, just like offcial package. Please have a try: https://github.com/xuhuisheng/rocm-gfx803/releases/download/rocm513/rocblas_2.43.0.50103-66-f0273f26.dirty_amd64.deb
And my RX580 is Polaris. before patched, there are still failed. https://github.com/ROCmSoftwarePlatform/rocBLAS/issues/1218
Hi, AMD seems updated ROCM 5.1.1 to build 50101 since a while so the dirty version is no longer considered superior and would get uninstalled with every next system update. So manual installation needs to be done each time.
It is also possible to pin the dirty version, but then all system updates are effectively blocked, as there are unmet dependencies:
Trying suggested way results in removal of entire ROCm.
I am not sure how APT handles the dependencies, but probably easiest way is to increase rocblas version to something like 2.43.0.99999 ? Btw latest ROCm is now 5.1.3, would it be possible to bump the version to this one, maybe that would work as well?