"src/turbomind/utils/gemm_test/xlnet_gemm_func.cc" did not exist on "9efcac38af58b7247e205c47efe090b4c6ec7574"
gen_linux.sh 4.23 KB
Newer Older
1
#!/bin/bash
2
# This script is intended to run inside the go generate
3
# working directory must be llm/generate/
4
5
6
7
8
9
10
11
12
13
14

# First we build our default built-in library which will be linked into the CGO
# binary as a normal dependency. This default build is CPU based.
#
# Then we build a CUDA dynamic library (although statically linked with the CUDA
# library dependencies for maximum portability)
#
# Then if we detect ROCm, we build a dynamically loaded ROCm lib.  ROCm is particularly
# important to be a dynamic lib even if it's the only GPU library detected because
# we can't redistribute the objectfiles but must rely on dynamic libraries at
# runtime, which could lead the server not to start if not present.
15
16
17
18

set -ex
set -o pipefail

19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
# See https://llvm.org/docs/AMDGPUUsage.html#processors for reference
amdGPUs() {
    GPU_LIST=(
        "gfx803"
        "gfx900"
        "gfx906:xnack-"
        "gfx908:xnack-"
        "gfx90a:xnack+"
        "gfx90a:xnack-"
        "gfx1010"
        "gfx1012"
        "gfx1030"
        "gfx1100"
        "gfx1101"
        "gfx1102"
    )
    (
        IFS=$';'
        echo "'${GPU_LIST[*]}'"
    )
}

41
echo "Starting linux generate script"
42
if [ -z "${CUDACXX}" -a -x /usr/local/cuda/bin/nvcc ]; then
43
44
    export CUDACXX=/usr/local/cuda/bin/nvcc
fi
45
COMMON_CMAKE_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off"
46
47
48
49
source $(dirname $0)/gen_common.sh
init_vars
git_module_setup
apply_patches
50

51
52
53
54
55
56
if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
    #
    # CPU first for the default library
    #
    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS}"
    BUILD_DIR="${LLAMACPP_DIR}/build/linux/cpu"
57

58
59
    build
    install
60

61
62
63
64
65
    # Placeholder to keep go embed happy until we start building dynamic CPU lib variants
    touch ${BUILD_DIR}/lib/dummy.so
else
    echo "Skipping CPU generation step as requested"
fi
Daniel Hiltgen's avatar
Daniel Hiltgen committed
66

67
68
69
if [ -d /usr/local/cuda/lib64/ ]; then
    echo "CUDA libraries detected - building dynamic CUDA library"
    init_vars
70
71
72
73
    CUDA_MAJOR=$(ls /usr/local/cuda/lib64/libcudart.so.* | head -1 | cut -f3 -d. || true)
    if [ -n "${CUDA_MAJOR}" ]; then
        CUDA_VARIANT=_v${CUDA_MAJOR}
    fi
74
    CMAKE_DEFS="-DLLAMA_CUBLAS=on ${COMMON_CMAKE_DEFS} ${CMAKE_DEFS}"
75
    BUILD_DIR="${LLAMACPP_DIR}/build/linux/cuda${CUDA_VARIANT}"
76
77
    CUDA_LIB_DIR=/usr/local/cuda/lib64
    build
78
79
    install
    gcc -fPIC -g -shared -o ${BUILD_DIR}/lib/libext_server.so \
80
        -Wl,--whole-archive \
81
82
83
        ${BUILD_DIR}/lib/libext_server.a \
        ${BUILD_DIR}/lib/libcommon.a \
        ${BUILD_DIR}/lib/libllama.a \
84
85
86
87
88
89
90
        -Wl,--no-whole-archive \
        ${CUDA_LIB_DIR}/libcudart_static.a \
        ${CUDA_LIB_DIR}/libcublas_static.a \
        ${CUDA_LIB_DIR}/libcublasLt_static.a \
        ${CUDA_LIB_DIR}/libcudadevrt.a \
        ${CUDA_LIB_DIR}/libculibos.a \
        -lrt -lpthread -ldl -lstdc++ -lm
91
fi
92

93
if [ -z "${ROCM_PATH}" ]; then
94
95
96
97
    # Try the default location in case it exists
    ROCM_PATH=/opt/rocm
fi

98
if [ -z "${CLBlast_DIR}" ]; then
99
100
101
102
103
104
    # Try the default location in case it exists
    if [ -d /usr/lib/cmake/CLBlast ]; then
        export CLBlast_DIR=/usr/lib/cmake/CLBlast
    fi
fi

105
106
if [ -d "${ROCM_PATH}" ]; then
    echo "ROCm libraries detected - building dynamic ROCm library"
107
108
109
    if [ -f ${ROCM_PATH}/lib/librocm_smi64.so.? ]; then
        ROCM_VARIANT=_v$(ls ${ROCM_PATH}/lib/librocm_smi64.so.? | cut -f3 -d. || true)
    fi
110
    init_vars
111
    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DLLAMA_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
112
    BUILD_DIR="${LLAMACPP_DIR}/build/linux/rocm${ROCM_VARIANT}"
113
    build
114
115
    install
    gcc -fPIC -g -shared -o ${BUILD_DIR}/lib/libext_server.so \
116
        -Wl,--whole-archive \
117
118
119
        ${BUILD_DIR}/lib/libext_server.a \
        ${BUILD_DIR}/lib/libcommon.a \
        ${BUILD_DIR}/lib/libllama.a \
120
121
122
123
124
125
        -Wl,--no-whole-archive \
        -lrt -lpthread -ldl -lstdc++ -lm \
        -L/opt/rocm/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ \
        -Wl,-rpath,/opt/rocm/lib,-rpath,/opt/amdgpu/lib/x86_64-linux-gnu/ \
        -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu
fi
126
127

cleanup