diff --git a/llm/generate/gen_common.sh b/llm/generate/gen_common.sh index 16ff710a..da1b0688 100644 --- a/llm/generate/gen_common.sh +++ b/llm/generate/gen_common.sh @@ -21,7 +21,7 @@ init_vars() { # TODO - add additional optimization flags... CMAKE_DEFS="-DCMAKE_BUILD_TYPE=Release -DLLAMA_SERVER_VERBOSE=off ${CMAKE_DEFS}" fi - case $(uname -s) in + case $(uname -s) in "Darwin") LIB_EXT="dylib" WHOLE_ARCHIVE="-Wl,-force_load" diff --git a/llm/generate/gen_linux.sh b/llm/generate/gen_linux.sh index b402ecad..63668bd2 100755 --- a/llm/generate/gen_linux.sh +++ b/llm/generate/gen_linux.sh @@ -165,11 +165,11 @@ if [ -d "${CUDA_LIB_DIR}" ]; then fi if [ "${ARCH}" == "arm64" ]; then echo "ARM CPU detected - disabling unsupported AVX instructions" - + # ARM-based CPUs such as M1 and Tegra do not support AVX extensions. # - # CUDA compute < 6.0 lacks proper FP16 support on ARM. - # Disabling has minimal performance effect while maintaining compatibility. + # CUDA compute < 6.0 lacks proper FP16 support on ARM. + # Disabling has minimal performance effect while maintaining compatibility. ARM64_DEFS="-DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_CUDA_F16=off" fi # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp