baalajimaestro/ollama
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

Add compute capability 5.0, 7.5, and 8.0

Browse source
This commit is contained in:
Daniel Hiltgen 2024-01-20 12:15:50 -08:00
parent 681a914990
commit a447a083f2
5 changed files with 14 additions and 6 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

3
docs/development.md
View file

@ -50,7 +50,8 @@ development and runtime packages.
Typically the build scripts will auto-detect CUDA, however, if your Linux distro Typically the build scripts will auto-detect CUDA, however, if your Linux distro
or installation approach uses unusual paths, you can specify the location by or installation approach uses unusual paths, you can specify the location by
specifying an environment variable `CUDA_LIB_DIR` to the location of the shared specifying an environment variable `CUDA_LIB_DIR` to the location of the shared
libraries, and `CUDACXX` to the location of the nvcc compiler. libraries, and `CUDACXX` to the location of the nvcc compiler. You can customize
set set of target CUDA architectues by setting `CMAKE_CUDA_ARCHITECTURES` (e.g. "50;60;70")
Then generate dependencies: Then generate dependencies:

5
gpu/gpu.go
View file

@ -29,9 +29,8 @@ type handles struct {
var gpuMutex sync.Mutex var gpuMutex sync.Mutex
var gpuHandles *handles = nil var gpuHandles *handles = nil
// With our current CUDA compile flags, older than 5.2 will not work properly // With our current CUDA compile flags, older than 5.0 will not work properly
// 5.0: CUDA error: no kernel image is available for execution on the device var CudaComputeMin = [2]C.int{5, 0}
var CudaComputeMin = [2]C.int{5, 2}
// Possible locations for the nvidia-ml library // Possible locations for the nvidia-ml library
var CudaLinuxGlobs = []string{ var CudaLinuxGlobs = []string{

3
llm/generate/gen_common.sh
View file

@ -39,6 +39,9 @@ init_vars() {
*) *)
;; ;;
esac esac
if [ -z "${CMAKE_CUDA_ARCHITECTURES}" ] ; then
CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
fi
} }
git_module_setup() { git_module_setup() {

2
llm/generate/gen_linux.sh
View file

@ -125,7 +125,7 @@ if [ -d "${CUDA_LIB_DIR}" ]; then
if [ -n "${CUDA_MAJOR}" ]; then if [ -n "${CUDA_MAJOR}" ]; then
CUDA_VARIANT=_v${CUDA_MAJOR} CUDA_VARIANT=_v${CUDA_MAJOR}
fi fi
CMAKE_DEFS="-DLLAMA_CUBLAS=on -DLLAMA_CUDA_FORCE_MMQ=on ${COMMON_CMAKE_DEFS} ${CMAKE_DEFS}" CMAKE_DEFS="-DLLAMA_CUBLAS=on -DLLAMA_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${COMMON_CMAKE_DEFS} ${CMAKE_DEFS}"
BUILD_DIR="${LLAMACPP_DIR}/build/linux/${ARCH}/cuda${CUDA_VARIANT}" BUILD_DIR="${LLAMACPP_DIR}/build/linux/${ARCH}/cuda${CUDA_VARIANT}"
EXTRA_LIBS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda" EXTRA_LIBS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
build build

7
llm/generate/gen_windows.ps1
View file

@ -25,6 +25,11 @@ function init_vars {
} }
$script:GZIP=(get-command -ea 'silentlycontinue' gzip).path $script:GZIP=(get-command -ea 'silentlycontinue' gzip).path
$script:DUMPBIN=(get-command -ea 'silentlycontinue' dumpbin).path $script:DUMPBIN=(get-command -ea 'silentlycontinue' dumpbin).path
if ($null -eq $env:CMAKE_CUDA_ARCHITECTURES) {
$script:CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
} else {
$script:CMAKE_CUDA_ARCHITECTURES=$env:CMAKE_CUDA_ARCHITECTURES
}
} }
function git_module_setup { function git_module_setup {
@ -128,7 +133,7 @@ if ($null -ne $script:CUDA_LIB_DIR) {
} }
init_vars init_vars
$script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/cuda$script:CUDA_VARIANT" $script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/cuda$script:CUDA_VARIANT"
$script:cmakeDefs += @("-DLLAMA_CUBLAS=ON", "-DLLAMA_AVX=on") $script:cmakeDefs += @("-DLLAMA_CUBLAS=ON", "-DLLAMA_AVX=on", "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}")
build build
install install
cp "${script:CUDA_LIB_DIR}/cudart64_*.dll" "${script:buildDir}/lib" cp "${script:CUDA_LIB_DIR}/cudart64_*.dll" "${script:buildDir}/lib"