missing libs:
- libnccl.so.2
- libcublas.so.12
- libcublasLt.so.12
- libnvblas.so.12 (I think it asked this one too, not sure now, log is gone..)
- libcudart.so.12
I downloaded these and placed them all (with above symlinks) at cuda-12.8 folder:
- libnccl.so.2.26.2
- libcublas.so.12.8.5.7
- libcublasLt.so.12.8.5.7
- libnvblas.so.12.8.5.7
- libcudart.so.12.8.90
I downloaded these as they were the newest 12.8 that contain the missing .12 files
https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-cudart-12-8_12.8.90-1_amd64.deb
https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/libcublas12-cuda-12_12.8.5.7-1_amd64.deb
https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/libnccl2_2.26.2-1+cuda12.8_amd64.deb
(beware to not download any -dev package if they are huge! they have not .so files!!!! ask google AI like "show files for this ")
I wonder if I should have downloaded the max 12 version (12.9..)?
anyway, now I can run llama-server from
https://github.com/ai-dock/llama.cpp-cuda/releases/download/b9371/llama.cpp-b9371-cuda-12.8-amd64.tar.gz
with Llama-3.2-1B-Instruct-IQ3_M.gguf just to test it
I wonder if I could have done it in a better way?
as I did just guess work (just looked for 12.8 max available).
It was not complicated, but I am not sure it is working as expected, I mean at best possible as I just guessed the dependencies.
Should I have installed some repo? but I dont want to break other things thru package conflicts as ubuntu is complicated (for me) to recover.
missing libs:
I downloaded these and placed them all (with above symlinks) at cuda-12.8 folder:
I downloaded these as they were the newest 12.8 that contain the missing .12 files
https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-cudart-12-8_12.8.90-1_amd64.deb
https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/libcublas12-cuda-12_12.8.5.7-1_amd64.deb
https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/libnccl2_2.26.2-1+cuda12.8_amd64.deb
(beware to not download any -dev package if they are huge! they have not .so files!!!! ask google AI like "show files for this ")
I wonder if I should have downloaded the max 12 version (12.9..)?
anyway, now I can run llama-server from
https://github.com/ai-dock/llama.cpp-cuda/releases/download/b9371/llama.cpp-b9371-cuda-12.8-amd64.tar.gz
with
Llama-3.2-1B-Instruct-IQ3_M.ggufjust to test itI wonder if I could have done it in a better way?
as I did just guess work (just looked for 12.8 max available).
It was not complicated, but I am not sure it is working as expected, I mean at best possible as I just guessed the dependencies.
Should I have installed some repo? but I dont want to break other things thru package conflicts as ubuntu is complicated (for me) to recover.