diff --git a/.github/workflows/release-prism.yml b/.github/workflows/release-prism.yml index ff4d09584d1e..3945c94a731b 100644 --- a/.github/workflows/release-prism.yml +++ b/.github/workflows/release-prism.yml @@ -178,6 +178,8 @@ jobs: -DCMAKE_INSTALL_RPATH='$ORIGIN' \ -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \ -DGGML_NATIVE=OFF \ + -DGGML_BACKEND_DL=ON \ + -DGGML_CPU_ALL_VARIANTS=ON \ -DGGML_CUDA=ON \ ${{ env.CMAKE_ARGS }} cmake --build build --config Release -j $(nproc) 2>&1 | grep -v "^nvcc warning" @@ -301,10 +303,14 @@ jobs: call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }} rem arm64: ggml-cpu refuses MSVC for ARM (needs clang), so that leg builds only the rem ggml-cuda.dll companion for the BACKEND_DL cpu-arm64 zip, like upstream does. + rem x64: BACKEND_DL + CPU_ALL_VARIANTS so the bundled CPU backend is picked at runtime. + rem GGML_NATIVE=OFF alone compiles one ggml-cpu.dll with AVX2/FMA, which dies silently + rem at startup on CPUs without AVX2 (#241). cmake -S . -B build -G "Ninja Multi-Config" ^ -DGGML_NATIVE=OFF ^ -DGGML_CUDA=ON ^ - ${{ matrix.arch == 'arm64' && '-DGGML_BACKEND_DL=ON -DGGML_CPU=OFF' || '' }} ^ + -DGGML_BACKEND_DL=ON ^ + ${{ matrix.arch == 'arm64' && '-DGGML_CPU=OFF' || '-DGGML_CPU_ALL_VARIANTS=ON' }} ^ -DLLAMA_BUILD_BORINGSSL=ON ^ -DCMAKE_CUDA_FLAGS="-diag-suppress=221" ^ ${{ matrix.defines }} ${{ env.CMAKE_ARGS }} @@ -623,6 +629,7 @@ jobs: -DCMAKE_BUILD_TYPE=Release ` -DGGML_BACKEND_DL=ON ` -DGGML_NATIVE=OFF ` + -DGGML_CPU_ALL_VARIANTS=ON ` -DGGML_OPENMP=OFF ` -DGPU_TARGETS="${{ matrix.gpu_targets }}" ` -DGGML_HIP_ROCWMMA_FATTN=ON `