LostRuins / koboldcpp

A simple one-file way to run various GGML and GGUF models with KoboldAI's UI
https://github.com/lostruins/koboldcpp
GNU Affero General Public License v3.0
4.32k stars 310 forks source link

rocBLAS error: Cannot read TensileLibrary.dat: No such file or directory for GPU arch : gfx1036 #676

Open f1am3d opened 4 months ago

f1am3d commented 4 months ago

Error on loading gguf model:

rocBLAS error: Cannot read \AppData\Local\Temp\_MEI59482\/library/TensileLibrary.dat: No such file or directory for GPU arch : gfx1036

YellowRoseCx commented 4 months ago

I don't think iGPUs are supported on ROCm for Windows or Linux

f1am3d commented 4 months ago

I don't think iGPUs are supported on ROCm for Windows or Linux

RX 7900 XTX is not an integrated GPU

YellowRoseCx commented 4 months ago

I don't think iGPUs are supported on ROCm for Windows or Linux

RX 7900 XTX is not an integrated GPU

then that likely means you have an AMD CPU with integrated graphics that haven't been disabled in the BIOS. AMD's ROCm install documents says "The iGPU is non-essential for AI and ML workloads and not officially supported. Disable iGPU in SBIOS before proceeding to avoid unknown issues." https://rocm.docs.amd.com/projects/radeon/en/latest/docs/prerequisites.html#disable-igpu Though, it's buried and not in the Windows side of the documents so I don't blame you if you didn't see it

brknsoul commented 4 months ago

I'm getting this too, for my 6600xt

rocBLAS error: Cannot read D:\WinTemp\_MEI96962\/library/TensileLibrary.dat: No such file or directory for GPU arch : gfx1032
 List of available TensileLibrary Files :

Using the new release of koboldcpp_rocm I just installed ROCm HIP SDK for windows.. i'll reboot to see if that changes things.

EDIT: Chatted with you on discord.. I understand a new build is en route. Thank you!

f1am3d commented 4 months ago

then that likely means you have an AMD CPU with integrated graphics that haven't been disabled in the BIOS. AMD's ROCm install documents says "The iGPU is non-essential for AI and ML workloads and not officially supported. Disable iGPU in SBIOS before proceeding to avoid unknown issues." https://rocm.docs.amd.com/projects/radeon/en/latest/docs/prerequisites.html#disable-igpu Though, it's buried and not in the Windows side of the documents so I don't blame you if you didn't see it

Okay, I will try to disable it. But also from this list I noted that exe is compiled only for 6xxx series of GPUs, why there are no 7xxx? Is it a problem to compile for 7th generation?

https://github.com/YellowRoseCx/koboldcpp-rocm/releases/tag/v1.57.1.yr1-ROCm

jasyuiop commented 4 months ago

I'm getting this too, for my 6600xt

rocBLAS error: Cannot read D:\WinTemp\_MEI96962\/library/TensileLibrary.dat: No such file or directory for GPU arch : gfx1032
 List of available TensileLibrary Files :

Using the new release of koboldcpp_rocm I just installed ROCm HIP SDK for windows.. i'll reboot to see if that changes things.

EDIT: Chatted with you on discord.. I understand a new build is en route. Thank you!

@brknsoul Even if you install the HIP SDK and run koboldcpp without any problems, llm will start to bullshit and delusional.

https://github.com/LostRuins/koboldcpp/issues/655#issuecomment-1937510035 here I explained why, if @YellowRoseCx compiles it with the "none-lazy" kernel when he is available, we will have a solution that is much better than vulkan backend and clblast for gfx1032 cards for now.

YellowRoseCx commented 4 months ago

then that likely means you have an AMD CPU with integrated graphics that haven't been disabled in the BIOS. AMD's ROCm install documents says "The iGPU is non-essential for AI and ML workloads and not officially supported. Disable iGPU in SBIOS before proceeding to avoid unknown issues." https://rocm.docs.amd.com/projects/radeon/en/latest/docs/prerequisites.html#disable-igpu Though, it's buried and not in the Windows side of the documents so I don't blame you if you didn't see it

Okay, I will try to disable it. But also from this list I noted that exe is compiled only for 6xxx series of GPUs, why there are no 7xxx? Is it a problem to compile for 7th generation?

https://github.com/YellowRoseCx/koboldcpp-rocm/releases/tag/v1.57.1.yr1-ROCm

That list is just cards that were added, not a complete list of working cards