Bug: llama-server-cuda docker image build failure after #10320

[henryclw]

What happened?

After #10320, the docker image ci for llama-server-cuda failed.

https://github.com/ggerganov/llama.cpp/actions/runs/11877198432/job/33096358920

This bug could be reproduced locally.

Name and Version

After Commit 467576b, docker ci failed.

What operating system are you seeing the problem on?

No response

Relevant log output

#13 [8/8] RUN if [ "default" != "default" ]; then         export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=default";     fi &&     cmake -B build -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . &&     cmake --build build --config Release -j$(nproc) &&     cp build/bin/* .
#13 0.130 -- The C compiler identification is GNU 11.4.0
#13 0.182 -- The CXX compiler identification is GNU 11.4.0
#13 0.189 -- Detecting C compiler ABI info
#13 0.253 -- Detecting C compiler ABI info - done
#13 0.261 -- Check for working C compiler: /usr/bin/cc - skipped
#13 0.262 -- Detecting C compile features
#13 0.262 -- Detecting C compile features - done
#13 0.265 -- Detecting CXX compiler ABI info
#13 0.338 -- Detecting CXX compiler ABI info - done
#13 0.346 -- Check for working CXX compiler: /usr/bin/c++ - skipped
#13 0.346 -- Detecting CXX compile features
#13 0.347 -- Detecting CXX compile features - done
#13 0.351 -- Found Git: /usr/bin/git (found version "2.34.1") 
#13 0.496 -- Looking for pthread.h
#13 0.565 -- Looking for pthread.h - found
#13 0.565 -- Performing Test CMAKE_HAVE_LIBC_PTHREAD
#13 0.633 -- Performing Test CMAKE_HAVE_LIBC_PTHREAD - Success
#13 0.634 -- Found Threads: TRUE  
#13 0.636 -- Warning: ccache not found - consider installing it for faster compilation or disable this warning with GGML_CCACHE=OFF
#13 0.653 -- CMAKE_SYSTEM_PROCESSOR: x86_64
#13 0.876 -- Found OpenMP_C: -fopenmp (found version "4.5") 
#13 0.957 -- Found OpenMP_CXX: -fopenmp (found version "4.5") 
#13 0.958 -- Found OpenMP: TRUE (found version "4.5")  
#13 0.958 -- OpenMP found
#13 0.958 -- Using llamafile
#13 0.959 -- x86 detected
#13 0.959 -- Using runtime weight conversion of Q4_0 to Q4_0_x_x to enable optimized GEMM/GEMV kernels
#13 0.959 -- Including CPU backend
#13 0.960 -- Using AMX
#13 0.960 -- Including AMX backend
#13 0.976 -- Found CUDAToolkit: /usr/local/cuda/include (found version "12.6.20") 
#13 0.997 -- CUDA Toolkit found
#13 0.997 -- Using CUDA architectures: native
#13 1.019 CMake Error at /usr/share/cmake-3.22/Modules/CMakeDetermineCompilerId.cmake:726 (message):
#13 1.019   Compiling the CUDA compiler identification source file
#13 1.019   "CMakeCUDACompilerId.cu" failed.
#13 1.019 
#13 1.019   Compiler: /usr/local/cuda/bin/nvcc
#13 1.019 
#13 1.019   Build flags:
#13 1.019 
#13 1.019   Id flags: --keep;--keep-dir;tmp;-gencode=arch=compute_,code=sm_ -v
#13 1.019 
#13 1.019   
#13 1.019 
#13 1.019   The output was:
#13 1.019 
#13 1.019   1
#13 1.019 
#13 1.019   nvcc fatal : Unsupported gpu architecture 'compute_'
#13 1.019 
#13 1.019   
#13 1.019 
#13 1.019   
#13 1.019 
#13 1.019 Call Stack (most recent call first):
#13 1.019   /usr/share/cmake-3.22/Modules/CMakeDetermineCompilerId.cmake:6 (CMAKE_DETERMINE_COMPILER_ID_BUILD)
#13 1.019   /usr/share/cmake-3.22/Modules/CMakeDetermineCompilerId.cmake:48 (__determine_compiler_id_test)
#13 1.019   /usr/share/cmake-3.22/Modules/CMakeDetermineCUDACompiler.cmake:298 (CMAKE_DETERMINE_COMPILER_ID)
#13 1.019   ggml/src/ggml-cuda/CMakeLists.txt:25 (enable_language)
#13 1.019 
#13 1.019 
#13 1.020 -- Configuring incomplete, errors occurred!
#13 1.020 See also "/app/build/CMakeFiles/CMakeOutput.log".
#13 1.020 See also "/app/build/CMakeFiles/CMakeError.log".
#13 ERROR: process "/bin/sh -c if [ \"${CUDA_DOCKER_ARCH}\" != \"default\" ]; then         export CMAKE_ARGS=\"-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}\";     fi &&     cmake -B build -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . &&     cmake --build build --config Release -j$(nproc) &&     cp build/bin/* ." did not complete successfully: exit code: 1