780m support

[Adriankhl]

Hi, I am running a laptop with 780m apu. I am curious about running the rocm version, but the gfx1103 support is not included in the compilation process so I cannot run hiprocm. I did attempt to build it myself but I failed because of ROCm/ROCm#2785. @YellowRoseCx How do you get around with the issue?

[YellowRoseCx]

Hi, I am running a laptop with 780m apu. I am curious about running the rocm version, but the gfx1103 support is not included in the compilation process so I cannot run hiprocm. I did attempt to build it myself but I failed because of ROCm/ROCm#2785. @YellowRoseCx How do you get around with the issue?

Are you on Windows or Linux? Because on Windows, I'm getting that same exact error which is why I haven't been able to build the latest version for Windows. But with Linux using the makefile, it works just fine doing make LLAMA_HIPBLAS=1 -j4

If you're on Linux, try using the command HSA_OVERRIDE_GFX_VERSION=11.0.0 before running koboldcpp.py after compiling, But if you're on Windows, then I believe the only way it could possibly work as if you were to compile the TensileLibraries yourself, but even then, it might not be supported.

[Adriankhl]

Thanks for the information. This is a windows machine because I also want to experiment with directml. So I probably need to wait for a better support from AMD (which may never come, I will keep checking ROCm/ROCm#2631). In the meanwhile I think the vulkan backend is very interesting.