sci-misc/llama-cpp build failure when enabling USE flag "blas" for sci-libs/ggml

[fpemud]

emerge command (I've limited the USE flags to the bare minimum. The error still persists when other USE flags are enabled):

USE="blas -cuda -kompute -mkl -opencl -openmp -rocm -sycl -vulkan" emerge -1 ggml llama-cpp whisper-cpp

ggml installs successfully, but error occurs when installing llama-cpp and whisper-cpp. Error:

/usr/lib/gcc/x86_64-pc-linux-gnu/14/../../../../x86_64-pc-linux-gnu/bin/ld: /usr/lib64/libggml-blas.so: undefined reference to `cblas_sgemm'

I think full build log is not needed.
(in fact I have problem uploading it due to my network)

I know that sci-libs/ggml depends on sci-libs/openblas from the ebuild file, USE flags of sci-libs/openblas:

sh-5.2# equery uses openblas
[ Legend : U - final flag setting for installation]
[        : I - package is installed with flag     ]
[ Colors : set, unset                             ]
 * Found these USE flags for sci-libs/openblas-0.3.26-r1:
 U I
 - - dynamic      : Build dynamic architecture detection at run time (for multi targets) 
 + + eselect-ldso : Enable BLAS/LAPACK runtime switching 
 - - index-64bit  : Compile a separate INTERFACE64 OpenBLAS 
 + + openmp       : Build support for the OpenMP (support parallel computing), requires >=sys-devel/gcc-4.2 built with USE="openmp"
 - - pthread      : Build with pthread threading model 
 - - relapack     : Build the ReLAPACK extension to OpenBLAS 
 - - test         : Enable dependencies and/or preparations necessary to run tests (usually controlled by FEATURES=test but can be toggled
                    independently)

[longlene]

@fpemud let me take a look, never tried the minimal build before

[longlene]

@fpemud It should be good now.

[fpemud]

The following command runs successfully:

USE="blas -cuda -kompute -mkl -opencl -openmp -rocm -sycl -vulkan" emerge -1 ggml llama-cpp whisper-cpp

But this error comes back when "rocm" USE flag is enabled:

-- The HIP compiler identification is Clang 18.1.8
CMake Error at /usr/share/cmake/Modules/CMakeDetermineHIPCompiler.cmake:217 (message):
  The ROCm root directory:

   /usr/local

  does not contain the HIP runtime CMake package, expected at one of:

   /usr/local/lib/cmake/hip-lang/hip-lang-config.cmake
   /usr/local/lib64/cmake/hip-lang/hip-lang-config.cmake

Call Stack (most recent call first):
  src/ggml-hip/CMakeLists.txt:36 (enable_language)

I see you reverted this line in ggml-0_p20250204.ebuild:

-DGGML_HIP=$(usex rocm)

I don't know the detail, but I'm also feeling that GGML_HIP is better than GGML_HIPBLAS.
Can you tell me what do you think about this error?
I'm curious why this error happens. Why does cmake go to /usr/local to find files?

[longlene]

@fpemud It may be a mistake, let me check later, I changed it just based on the cmake files.

[longlene]

@fpemud I changed it from HIPBLAS TO HIP, because I can't find any HIPBLAS word in cmake, just add it back.

[fpemud]

I managed to build sci-libs/ggml with rocm USE flag after adding "-DCMAKE_HIP_COMPILER_ROCM_ROOT=/usr" in "mycmakeargs".
This time I confirmed that my GPU is actually being used by llama-cpp (but the tps is still very low, I'm still on it).
I don't think this is the final solution.
I guess this is due to /usr/share/cmake/Modules/CMakeDetermineHIPCompiler.cmake from dev-build/cmake.