2023-11-29 19:00:37 +00:00
|
|
|
#!/bin/bash
|
2024-04-09 22:57:45 +00:00
|
|
|
# This script is intended to run inside the go generate
|
|
|
|
# working directory must be llm/generate/
|
2023-12-20 18:36:01 +00:00
|
|
|
|
2024-01-13 00:28:00 +00:00
|
|
|
# First we build one or more CPU based LLM libraries
|
2023-12-20 18:36:01 +00:00
|
|
|
#
|
2024-01-13 00:28:00 +00:00
|
|
|
# Then if we detect CUDA, we build a CUDA dynamic library, and carry the required
|
|
|
|
# library dependencies
|
2023-12-20 18:36:01 +00:00
|
|
|
#
|
2024-01-13 00:28:00 +00:00
|
|
|
# Then if we detect ROCm, we build a dynamically loaded ROCm lib. The ROCM
|
|
|
|
# libraries are quite large, and also dynamically load data files at runtime
|
|
|
|
# which in turn are large, so we don't attempt to cary them as payload
|
2023-11-14 01:20:34 +00:00
|
|
|
|
|
|
|
set -ex
|
|
|
|
set -o pipefail
|
|
|
|
|
2024-01-03 23:12:29 +00:00
|
|
|
# See https://llvm.org/docs/AMDGPUUsage.html#processors for reference
|
|
|
|
amdGPUs() {
|
2024-01-21 20:57:13 +00:00
|
|
|
if [ -n "${AMDGPU_TARGETS}" ]; then
|
|
|
|
echo "${AMDGPU_TARGETS}"
|
|
|
|
return
|
|
|
|
fi
|
2024-01-03 23:12:29 +00:00
|
|
|
GPU_LIST=(
|
|
|
|
"gfx900"
|
|
|
|
"gfx906:xnack-"
|
|
|
|
"gfx908:xnack-"
|
|
|
|
"gfx90a:xnack+"
|
|
|
|
"gfx90a:xnack-"
|
2024-03-15 22:34:58 +00:00
|
|
|
"gfx940"
|
|
|
|
"gfx941"
|
|
|
|
"gfx942"
|
2024-01-03 23:12:29 +00:00
|
|
|
"gfx1010"
|
|
|
|
"gfx1012"
|
|
|
|
"gfx1030"
|
|
|
|
"gfx1100"
|
|
|
|
"gfx1101"
|
|
|
|
"gfx1102"
|
|
|
|
)
|
|
|
|
(
|
|
|
|
IFS=$';'
|
|
|
|
echo "'${GPU_LIST[*]}'"
|
|
|
|
)
|
|
|
|
}
|
|
|
|
|
2023-11-14 01:20:34 +00:00
|
|
|
echo "Starting linux generate script"
|
2024-01-09 20:55:36 +00:00
|
|
|
if [ -z "${CUDACXX}" ]; then
|
|
|
|
if [ -x /usr/local/cuda/bin/nvcc ]; then
|
|
|
|
export CUDACXX=/usr/local/cuda/bin/nvcc
|
|
|
|
else
|
|
|
|
# Try the default location in case it exists
|
|
|
|
export CUDACXX=$(command -v nvcc)
|
|
|
|
fi
|
2023-11-29 19:00:37 +00:00
|
|
|
fi
|
2024-07-06 01:48:31 +00:00
|
|
|
COMMON_CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_OPENMP=off"
|
2023-11-14 01:20:34 +00:00
|
|
|
source $(dirname $0)/gen_common.sh
|
|
|
|
init_vars
|
|
|
|
git_module_setup
|
|
|
|
apply_patches
|
2023-12-20 18:36:01 +00:00
|
|
|
|
2024-03-14 17:24:13 +00:00
|
|
|
init_vars
|
2024-04-17 18:43:28 +00:00
|
|
|
if [ -z "${OLLAMA_SKIP_STATIC_GENERATE}" -o "${OLLAMA_CPU_TARGET}" = "static" ]; then
|
|
|
|
# Builds by default, allows skipping, forces build if OLLAMA_CPU_TARGET="static"
|
|
|
|
# Enables optimized Dockerfile builds using a blanket skip and targeted overrides
|
|
|
|
# Static build for linking into the Go binary
|
|
|
|
init_vars
|
|
|
|
CMAKE_TARGETS="--target llama --target ggml"
|
2024-07-05 17:25:58 +00:00
|
|
|
CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DGGML_NATIVE=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_OPENMP=off ${CMAKE_DEFS}"
|
2024-04-17 18:43:28 +00:00
|
|
|
BUILD_DIR="../build/linux/${ARCH}_static"
|
|
|
|
echo "Building static library"
|
|
|
|
build
|
2024-04-17 17:04:28 +00:00
|
|
|
fi
|
2024-03-14 17:24:13 +00:00
|
|
|
|
2024-04-17 17:04:28 +00:00
|
|
|
init_vars
|
|
|
|
if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
|
2024-01-07 23:48:05 +00:00
|
|
|
# Users building from source can tune the exact flags we pass to cmake for configuring
|
|
|
|
# llama.cpp, and we'll build only 1 CPU variant in that case as the default.
|
|
|
|
if [ -n "${OLLAMA_CUSTOM_CPU_DEFS}" ]; then
|
2024-03-14 17:24:13 +00:00
|
|
|
init_vars
|
2024-01-07 23:48:05 +00:00
|
|
|
echo "OLLAMA_CUSTOM_CPU_DEFS=\"${OLLAMA_CUSTOM_CPU_DEFS}\""
|
2024-07-06 22:58:16 +00:00
|
|
|
CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
|
2024-03-14 17:24:13 +00:00
|
|
|
BUILD_DIR="../build/linux/${ARCH}/cpu"
|
2024-01-07 23:48:05 +00:00
|
|
|
echo "Building custom CPU"
|
|
|
|
build
|
2024-03-14 17:24:13 +00:00
|
|
|
compress
|
2024-01-07 23:48:05 +00:00
|
|
|
else
|
|
|
|
# Darwin Rosetta x86 emulation does NOT support AVX, AVX2, AVX512
|
2024-07-05 17:25:58 +00:00
|
|
|
# -DGGML_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
|
|
|
|
# -DGGML_F16C -- 2012 Intel Ivy Bridge & AMD 2011 Bulldozer (No significant improvement over just AVX)
|
|
|
|
# -DGGML_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
|
|
|
|
# -DGGML_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
|
2024-01-07 23:48:05 +00:00
|
|
|
# Note: the following seem to yield slower results than AVX2 - ymmv
|
2024-07-05 17:25:58 +00:00
|
|
|
# -DGGML_AVX512 -- 2017 Intel Skylake and High End DeskTop (HEDT)
|
|
|
|
# -DGGML_AVX512_VBMI -- 2018 Intel Cannon Lake
|
|
|
|
# -DGGML_AVX512_VNNI -- 2021 Intel Alder Lake
|
2023-12-23 19:35:44 +00:00
|
|
|
|
2024-07-06 22:58:16 +00:00
|
|
|
COMMON_CPU_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_OPENMP=off"
|
2024-01-21 20:57:13 +00:00
|
|
|
if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu" ]; then
|
|
|
|
#
|
|
|
|
# CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
|
|
|
|
#
|
2024-03-14 17:24:13 +00:00
|
|
|
init_vars
|
2024-07-05 17:25:58 +00:00
|
|
|
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
|
2024-03-14 17:24:13 +00:00
|
|
|
BUILD_DIR="../build/linux/${ARCH}/cpu"
|
2024-01-21 20:57:13 +00:00
|
|
|
echo "Building LCD CPU"
|
|
|
|
build
|
2024-03-14 17:24:13 +00:00
|
|
|
compress
|
2024-01-21 20:57:13 +00:00
|
|
|
fi
|
2023-12-14 01:26:47 +00:00
|
|
|
|
2024-03-25 15:07:44 +00:00
|
|
|
if [ "${ARCH}" == "x86_64" ]; then
|
2024-01-21 20:57:13 +00:00
|
|
|
#
|
2024-03-25 15:07:44 +00:00
|
|
|
# ARM chips in M1/M2/M3-based MACs and NVidia Tegra devices do not currently support avx extensions.
|
2024-01-21 20:57:13 +00:00
|
|
|
#
|
2024-03-25 15:07:44 +00:00
|
|
|
if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu_avx" ]; then
|
|
|
|
#
|
|
|
|
# ~2011 CPU Dynamic library with more capabilities turned on to optimize performance
|
|
|
|
# Approximately 400% faster than LCD on same CPU
|
|
|
|
#
|
|
|
|
init_vars
|
2024-07-05 17:25:58 +00:00
|
|
|
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
|
2024-03-14 17:24:13 +00:00
|
|
|
BUILD_DIR="../build/linux/${ARCH}/cpu_avx"
|
2024-03-25 15:07:44 +00:00
|
|
|
echo "Building AVX CPU"
|
|
|
|
build
|
2024-03-14 17:24:13 +00:00
|
|
|
compress
|
2024-03-25 15:07:44 +00:00
|
|
|
fi
|
2024-01-07 23:48:05 +00:00
|
|
|
|
2024-03-25 15:07:44 +00:00
|
|
|
if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu_avx2" ]; then
|
|
|
|
#
|
|
|
|
# ~2013 CPU Dynamic library
|
|
|
|
# Approximately 10% faster than AVX on same CPU
|
|
|
|
#
|
|
|
|
init_vars
|
2024-07-05 17:25:58 +00:00
|
|
|
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
|
2024-03-14 17:24:13 +00:00
|
|
|
BUILD_DIR="../build/linux/${ARCH}/cpu_avx2"
|
2024-03-25 15:07:44 +00:00
|
|
|
echo "Building AVX2 CPU"
|
|
|
|
build
|
2024-03-14 17:24:13 +00:00
|
|
|
compress
|
2024-03-25 15:07:44 +00:00
|
|
|
fi
|
2024-01-21 20:57:13 +00:00
|
|
|
fi
|
2024-01-07 23:48:05 +00:00
|
|
|
fi
|
2024-01-05 20:13:08 +00:00
|
|
|
else
|
|
|
|
echo "Skipping CPU generation step as requested"
|
|
|
|
fi
|
2024-01-04 00:08:34 +00:00
|
|
|
|
2024-01-17 11:50:11 +00:00
|
|
|
# If needed, look for the default CUDA toolkit location
|
|
|
|
if [ -z "${CUDA_LIB_DIR}" ] && [ -d /usr/local/cuda/lib64 ]; then
|
2024-01-09 20:55:36 +00:00
|
|
|
CUDA_LIB_DIR=/usr/local/cuda/lib64
|
|
|
|
fi
|
|
|
|
|
2024-01-17 11:50:11 +00:00
|
|
|
# If needed, look for CUDA on Arch Linux
|
|
|
|
if [ -z "${CUDA_LIB_DIR}" ] && [ -d /opt/cuda/targets/x86_64-linux/lib ]; then
|
|
|
|
CUDA_LIB_DIR=/opt/cuda/targets/x86_64-linux/lib
|
|
|
|
fi
|
|
|
|
|
2024-01-25 14:49:35 +00:00
|
|
|
# Allow override in case libcudart is in the wrong place
|
|
|
|
if [ -z "${CUDART_LIB_DIR}" ]; then
|
|
|
|
CUDART_LIB_DIR="${CUDA_LIB_DIR}"
|
|
|
|
fi
|
|
|
|
|
2024-05-15 22:56:43 +00:00
|
|
|
if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
|
2023-12-20 18:36:01 +00:00
|
|
|
echo "CUDA libraries detected - building dynamic CUDA library"
|
|
|
|
init_vars
|
2024-01-09 20:55:36 +00:00
|
|
|
CUDA_MAJOR=$(ls "${CUDA_LIB_DIR}"/libcudart.so.* | head -1 | cut -f3 -d. || true)
|
2024-01-05 20:13:08 +00:00
|
|
|
if [ -n "${CUDA_MAJOR}" ]; then
|
|
|
|
CUDA_VARIANT=_v${CUDA_MAJOR}
|
|
|
|
fi
|
2024-03-25 15:07:44 +00:00
|
|
|
if [ "${ARCH}" == "arm64" ]; then
|
|
|
|
echo "ARM CPU detected - disabling unsupported AVX instructions"
|
2024-04-25 18:32:26 +00:00
|
|
|
|
2024-03-25 15:07:44 +00:00
|
|
|
# ARM-based CPUs such as M1 and Tegra do not support AVX extensions.
|
|
|
|
#
|
2024-04-25 18:32:26 +00:00
|
|
|
# CUDA compute < 6.0 lacks proper FP16 support on ARM.
|
|
|
|
# Disabling has minimal performance effect while maintaining compatibility.
|
2024-07-05 17:25:58 +00:00
|
|
|
ARM64_DEFS="-DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_CUDA_F16=off"
|
2024-03-25 15:07:44 +00:00
|
|
|
fi
|
2024-04-17 20:00:48 +00:00
|
|
|
# Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
|
2024-04-18 23:18:10 +00:00
|
|
|
if [ -n "${OLLAMA_CUSTOM_CUDA_DEFS}" ]; then
|
|
|
|
echo "OLLAMA_CUSTOM_CUDA_DEFS=\"${OLLAMA_CUSTOM_CUDA_DEFS}\""
|
2024-07-05 17:25:58 +00:00
|
|
|
CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
|
2024-04-18 23:18:10 +00:00
|
|
|
echo "Building custom CUDA GPU"
|
2024-04-17 20:00:48 +00:00
|
|
|
else
|
2024-07-10 20:17:13 +00:00
|
|
|
CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_FLAGS=-t8 -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} -DCMAKE_LIBRARY_PATH=/usr/local/cuda/compat"
|
2024-04-17 20:00:48 +00:00
|
|
|
fi
|
|
|
|
CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS}"
|
2024-03-14 17:24:13 +00:00
|
|
|
BUILD_DIR="../build/linux/${ARCH}/cuda${CUDA_VARIANT}"
|
2024-01-13 00:28:00 +00:00
|
|
|
EXTRA_LIBS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
|
2023-12-20 18:36:01 +00:00
|
|
|
build
|
2024-01-13 00:28:00 +00:00
|
|
|
|
2024-03-25 15:07:44 +00:00
|
|
|
# Carry the CUDA libs as payloads to help reduce dependency burden on users
|
2024-01-13 00:28:00 +00:00
|
|
|
#
|
|
|
|
# TODO - in the future we may shift to packaging these separately and conditionally
|
|
|
|
# downloading them in the install script.
|
2024-03-14 17:24:13 +00:00
|
|
|
DEPS="$(ldd ${BUILD_DIR}/bin/ollama_llama_server )"
|
2024-01-13 00:28:00 +00:00
|
|
|
for lib in libcudart.so libcublas.so libcublasLt.so ; do
|
|
|
|
DEP=$(echo "${DEPS}" | grep ${lib} | cut -f1 -d' ' | xargs || true)
|
|
|
|
if [ -n "${DEP}" -a -e "${CUDA_LIB_DIR}/${DEP}" ]; then
|
2024-03-14 17:24:13 +00:00
|
|
|
cp "${CUDA_LIB_DIR}/${DEP}" "${BUILD_DIR}/bin/"
|
2024-01-13 00:28:00 +00:00
|
|
|
elif [ -e "${CUDA_LIB_DIR}/${lib}.${CUDA_MAJOR}" ]; then
|
2024-03-14 17:24:13 +00:00
|
|
|
cp "${CUDA_LIB_DIR}/${lib}.${CUDA_MAJOR}" "${BUILD_DIR}/bin/"
|
2024-01-25 14:49:35 +00:00
|
|
|
elif [ -e "${CUDART_LIB_DIR}/${lib}" ]; then
|
2024-03-14 17:24:13 +00:00
|
|
|
cp -d ${CUDART_LIB_DIR}/${lib}* "${BUILD_DIR}/bin/"
|
2024-01-13 00:28:00 +00:00
|
|
|
else
|
2024-03-14 17:24:13 +00:00
|
|
|
cp -d "${CUDA_LIB_DIR}/${lib}*" "${BUILD_DIR}/bin/"
|
2024-01-13 00:28:00 +00:00
|
|
|
fi
|
|
|
|
done
|
2024-03-14 17:24:13 +00:00
|
|
|
compress
|
2024-01-13 00:28:00 +00:00
|
|
|
|
2023-12-14 01:26:47 +00:00
|
|
|
fi
|
2023-11-29 19:00:37 +00:00
|
|
|
|
2024-05-24 03:18:27 +00:00
|
|
|
if [ -z "${ONEAPI_ROOT}" ]; then
|
|
|
|
# Try the default location in case it exists
|
|
|
|
ONEAPI_ROOT=/opt/intel/oneapi
|
|
|
|
fi
|
|
|
|
|
2024-06-07 15:32:49 +00:00
|
|
|
if [ -z "${OLLAMA_SKIP_ONEAPI_GENERATE}" -a -d "${ONEAPI_ROOT}" ]; then
|
2024-05-24 03:18:27 +00:00
|
|
|
echo "OneAPI libraries detected - building dynamic OneAPI library"
|
|
|
|
init_vars
|
|
|
|
source ${ONEAPI_ROOT}/setvars.sh --force # set up environment variables for oneAPI
|
|
|
|
CC=icx
|
2024-07-05 17:25:58 +00:00
|
|
|
CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL=ON -DGGML_SYCL_F16=OFF"
|
2024-05-24 03:18:27 +00:00
|
|
|
BUILD_DIR="../build/linux/${ARCH}/oneapi"
|
|
|
|
EXTRA_LIBS="-fsycl -Wl,-rpath,${ONEAPI_ROOT}/compiler/latest/lib,-rpath,${ONEAPI_ROOT}/mkl/latest/lib,-rpath,${ONEAPI_ROOT}/tbb/latest/lib,-rpath,${ONEAPI_ROOT}/compiler/latest/opt/oclfpga/linux64/lib -lOpenCL -lmkl_core -lmkl_sycl_blas -lmkl_intel_ilp64 -lmkl_tbb_thread -ltbb"
|
|
|
|
DEBUG_FLAGS="" # icx compiles with -O0 if we pass -g, so we must remove it
|
|
|
|
build
|
|
|
|
|
|
|
|
# copy oneAPI dependencies
|
|
|
|
for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -e sycl -e mkl -e tbb); do
|
|
|
|
cp "${dep}" "${BUILD_DIR}/bin/"
|
|
|
|
done
|
|
|
|
cp "${ONEAPI_ROOT}/compiler/latest/lib/libOpenCL.so" "${BUILD_DIR}/bin/"
|
|
|
|
cp "${ONEAPI_ROOT}/compiler/latest/lib/libimf.so" "${BUILD_DIR}/bin/"
|
|
|
|
cp "${ONEAPI_ROOT}/compiler/latest/lib/libintlc.so.5" "${BUILD_DIR}/bin/"
|
|
|
|
cp "${ONEAPI_ROOT}/compiler/latest/lib/libirng.so" "${BUILD_DIR}/bin/"
|
|
|
|
cp "${ONEAPI_ROOT}/compiler/latest/lib/libpi_level_zero.so" "${BUILD_DIR}/bin/"
|
|
|
|
cp "${ONEAPI_ROOT}/compiler/latest/lib/libsvml.so" "${BUILD_DIR}/bin/"
|
|
|
|
cp "${ONEAPI_ROOT}/compiler/latest/lib/libur_loader.so.0" "${BUILD_DIR}/bin/"
|
|
|
|
compress
|
|
|
|
fi
|
|
|
|
|
2023-12-20 18:36:01 +00:00
|
|
|
if [ -z "${ROCM_PATH}" ]; then
|
2023-11-29 19:00:37 +00:00
|
|
|
# Try the default location in case it exists
|
|
|
|
ROCM_PATH=/opt/rocm
|
|
|
|
fi
|
|
|
|
|
2023-12-20 18:36:01 +00:00
|
|
|
if [ -z "${CLBlast_DIR}" ]; then
|
2023-11-29 19:00:37 +00:00
|
|
|
# Try the default location in case it exists
|
|
|
|
if [ -d /usr/lib/cmake/CLBlast ]; then
|
|
|
|
export CLBlast_DIR=/usr/lib/cmake/CLBlast
|
|
|
|
fi
|
|
|
|
fi
|
|
|
|
|
2024-05-15 22:56:43 +00:00
|
|
|
if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
|
2023-12-20 18:36:01 +00:00
|
|
|
echo "ROCm libraries detected - building dynamic ROCm library"
|
2024-02-16 01:15:09 +00:00
|
|
|
if [ -f ${ROCM_PATH}/lib/librocblas.so.*.*.????? ]; then
|
|
|
|
ROCM_VARIANT=_v$(ls ${ROCM_PATH}/lib/librocblas.so.*.*.????? | cut -f5 -d. || true)
|
2024-01-05 20:13:08 +00:00
|
|
|
fi
|
2023-11-29 19:00:37 +00:00
|
|
|
init_vars
|
2024-07-05 19:46:28 +00:00
|
|
|
CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DGGML_HIPBLAS=on -DLLAMA_CUDA_NO_PEER_COPY=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
|
2024-04-17 20:00:48 +00:00
|
|
|
# Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
|
2024-04-18 23:18:10 +00:00
|
|
|
if [ -n "${OLLAMA_CUSTOM_ROCM_DEFS}" ]; then
|
|
|
|
echo "OLLAMA_CUSTOM_ROCM_DEFS=\"${OLLAMA_CUSTOM_ROCM_DEFS}\""
|
|
|
|
CMAKE_DEFS="${CMAKE_DEFS} ${OLLAMA_CUSTOM_ROCM_DEFS}"
|
|
|
|
echo "Building custom ROCM GPU"
|
2024-04-17 20:00:48 +00:00
|
|
|
fi
|
2024-03-14 17:24:13 +00:00
|
|
|
BUILD_DIR="../build/linux/${ARCH}/rocm${ROCM_VARIANT}"
|
2024-03-11 15:45:57 +00:00
|
|
|
EXTRA_LIBS="-L${ROCM_PATH}/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ -Wl,-rpath,\$ORIGIN/../../rocm/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
|
2023-11-29 19:00:37 +00:00
|
|
|
build
|
2024-01-13 00:28:00 +00:00
|
|
|
|
2024-02-16 01:15:09 +00:00
|
|
|
# Record the ROCM dependencies
|
2024-03-14 17:24:13 +00:00
|
|
|
rm -f "${BUILD_DIR}/bin/deps.txt"
|
|
|
|
touch "${BUILD_DIR}/bin/deps.txt"
|
|
|
|
for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -e rocm -e amdgpu -e libtinfo ); do
|
|
|
|
echo "${dep}" >> "${BUILD_DIR}/bin/deps.txt"
|
2024-02-16 01:15:09 +00:00
|
|
|
done
|
2024-03-10 21:45:38 +00:00
|
|
|
# bomb out if for some reason we didn't get a few deps
|
2024-03-14 17:24:13 +00:00
|
|
|
if [ $(cat "${BUILD_DIR}/bin/deps.txt" | wc -l ) -lt 8 ] ; then
|
|
|
|
cat "${BUILD_DIR}/bin/deps.txt"
|
2024-03-10 21:45:38 +00:00
|
|
|
echo "ERROR: deps file short"
|
|
|
|
exit 1
|
|
|
|
fi
|
2024-03-14 17:24:13 +00:00
|
|
|
compress
|
2023-11-29 19:00:37 +00:00
|
|
|
fi
|
2023-12-22 17:51:53 +00:00
|
|
|
|
|
|
|
cleanup
|
2024-04-09 22:57:45 +00:00
|
|
|
echo "go generate completed. LLM runners: $(cd ${BUILD_DIR}/..; echo *)"
|