cublas.h and cublas_v2.h are missing in /usr/local/cuda-11.2/include/ about studio-lab-examples HOT 6 CLOSED

Sure, since I’m out of office, I’ll rather write briefly, and will augment later if you find necessary.

Reproduction step

1. Clone llama.cpp
2. follows the instruction to build cublas supported version: https://github.com/ggerganov/llama.cpp#cublas
3. Then will see an error, basically the same one as the issue content of this, originated from https://github.com/ggerganov/llama.cpp/blob/master/CMakeLists.txt#L316

Rather simple step to reproduce

CMAKE_ARGS="-DLLAMA_CUBLAS=on" pip install llama-cpp-python

will trigger the exact steps as I described in above reproduction step section.

from studio-lab-examples.

Thank you for using Studio Lab! As error message CMakeLists.txt:358 suggests, could please refer the CMakeLists file?
If you find out this error is caused by Studio Lab environment, please raise a new issue.

from studio-lab-examples.

I suspect that

Unable to find cublas_v2.h in either "/usr/local/cuda/include" or "/usr/math_libs/include"

is the source of error.
I myself met this error also. @icoxfog417 can you open it? As this is on going issue.

from studio-lab-examples.

@HyunggyuJang We sincerely appreciate your effort in reporting the issue. Could you kindly provide us with the reproducible steps? It would be immensely helpful for us to gain a better understanding of the specific libraries or tools you are attempting to install. We would like to suggest considering the use of conda for package installation, installing by manual will cause unexpected errors.

from studio-lab-examples.

Thank you. Did you install cmake and cuda-toolkit (11.4) before installing llama.cpp? (conda-forge version seems not include cuBLAS and cuda-toolkit includes it).

from studio-lab-examples.

I believe I tried. Let me check it again ASAP!

Edit: Still is an issue, yet rather different error message,

      -- Found CUDAToolkit: /home/studio-lab-user/.conda/envs/textgen/include (found version "11.4.48")
      -- cuBLAS found
      CMake Error at /tmp/pip-build-env-bkzhlka1/overlay/lib/python3.10/site-packages/cmake/data/share/cmake-3.27/Modules/CMakeDetermineCompilerId.cmake:753 (message):
        Compiling the CUDA compiler identification source file
        "CMakeCUDACompilerId.cu" failed.
      
        Compiler: /home/studio-lab-user/.conda/envs/textgen/bin/nvcc
      
        Build flags:
      
        Id flags: --keep;--keep-dir;tmp -v
      
      
      
        The output was:
      
        1
      
        gcc: fatal error: cannot execute 'cc1plus': execvp: No such file or
        directory
      
        compilation terminated.
      
        nvcc fatal : Failed to preprocess host compiler properties.

I'd better to experiment with it from clean state later. As I've tried on custom conda env where python version is 3.10. I'll try with default (studiolab) env.

from studio-lab-examples.