Switch back to subprocessing for llama.cpp

This should resolve a number of memory leak and stability defects by allowing
us to isolate llama.cpp in a separate process and shutdown when idle, and
gracefully restart if it has problems.  This also serves as a first step to be
able to run multiple copies to support multiple models concurrently.
This commit is contained in:
Daniel Hiltgen
2024-03-14 10:24:13 -07:00
parent 3b6a9154dd
commit 58d95cc9bd
35 changed files with 1416 additions and 1910 deletions

View File

@@ -18,21 +18,31 @@ sign() {
fi
}
COMMON_DARWIN_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.0 -DCMAKE_SYSTEM_NAME=Darwin"
COMMON_DARWIN_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.0 -DCMAKE_SYSTEM_NAME=Darwin -DLLAMA_METAL_EMBED_LIBRARY=on"
case "${GOARCH}" in
"amd64")
COMMON_CPU_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=off -DLLAMA_NATIVE=off"
# Static build for linking into the Go binary
init_vars
CMAKE_TARGETS="--target llama --target ggml"
CMAKE_DEFS="${COMMON_CPU_DEFS} -DBUILD_SHARED_LIBS=off -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}_static"
echo "Building static library"
build
#
# CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="${LLAMACPP_DIR}/build/darwin/${ARCH}/cpu"
BUILD_DIR="../build/darwin/${ARCH}/cpu"
echo "Building LCD CPU"
build
sign ${LLAMACPP_DIR}/build/darwin/${ARCH}/cpu/lib/libext_server.dylib
compress_libs
sign ${BUILD_DIR}/lib/libext_server.dylib
compress
#
# ~2011 CPU Dynamic library with more capabilities turned on to optimize performance
@@ -40,11 +50,11 @@ case "${GOARCH}" in
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="${LLAMACPP_DIR}/build/darwin/${ARCH}/cpu_avx"
BUILD_DIR="../build/darwin/${ARCH}/cpu_avx"
echo "Building AVX CPU"
build
sign ${LLAMACPP_DIR}/build/darwin/${ARCH}/cpu_avx/lib/libext_server.dylib
compress_libs
sign ${BUILD_DIR}/lib/libext_server.dylib
compress
#
# ~2013 CPU Dynamic library
@@ -52,20 +62,30 @@ case "${GOARCH}" in
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=on -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
BUILD_DIR="${LLAMACPP_DIR}/build/darwin/${ARCH}/cpu_avx2"
BUILD_DIR="../build/darwin/${ARCH}/cpu_avx2"
echo "Building AVX2 CPU"
EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation"
build
sign ${LLAMACPP_DIR}/build/darwin/${ARCH}/cpu_avx2/lib/libext_server.dylib
compress_libs
sign ${BUILD_DIR}/lib/libext_server.dylib
compress
;;
"arm64")
# Static build for linking into the Go binary
init_vars
CMAKE_TARGETS="--target llama --target ggml"
CMAKE_DEFS="${COMMON_CPU_DEFS} -DBUILD_SHARED_LIBS=off -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}_static"
echo "Building static library"
build
init_vars
CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DLLAMA_METAL_EMBED_LIBRARY=on -DLLAMA_ACCELERATE=on -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=on ${CMAKE_DEFS}"
BUILD_DIR="${LLAMACPP_DIR}/build/darwin/${ARCH}/metal"
BUILD_DIR="../build/darwin/${ARCH}/metal"
EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders"
build
sign ${LLAMACPP_DIR}/build/darwin/${ARCH}/metal/lib/libext_server.dylib
compress_libs
sign ${BUILD_DIR}/lib/libext_server.dylib
compress
;;
*)
echo "GOARCH must be set"
@@ -75,3 +95,4 @@ case "${GOARCH}" in
esac
cleanup
echo "go generate completed. LLM runners: $(cd ${BUILD_DIR}/..; echo *)"