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

Add support for CUDA 5.2 cards

Browse source
This commit is contained in:
Daniel Hiltgen 2024-01-20 10:48:43 -08:00
parent 4c54f0ddeb
commit 681a914990
2 changed files with 5 additions and 4 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

7
gpu/gpu.go
View file

@ -29,8 +29,9 @@ 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, 5.2 and older will not work properly // With our current CUDA compile flags, older than 5.2 will not work properly
const CudaComputeMajorMin = 6 // 5.0: CUDA error: no kernel image is available for execution on the device
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{
@ -133,7 +134,7 @@ func GetGPUInfo() GpuInfo {
if cc.err != nil { if cc.err != nil {
slog.Info(fmt.Sprintf("error looking up CUDA GPU compute capability: %s", C.GoString(cc.err))) slog.Info(fmt.Sprintf("error looking up CUDA GPU compute capability: %s", C.GoString(cc.err)))
C.free(unsafe.Pointer(cc.err)) C.free(unsafe.Pointer(cc.err))
} else if cc.major >= CudaComputeMajorMin { } else if cc.major > CudaComputeMin[0] || (cc.major == CudaComputeMin[0] && cc.minor >= CudaComputeMin[1]) {
slog.Info(fmt.Sprintf("CUDA Compute Capability detected: %d.%d", cc.major, cc.minor)) slog.Info(fmt.Sprintf("CUDA Compute Capability detected: %d.%d", cc.major, cc.minor))
resp.Library = "cuda" resp.Library = "cuda"
} else { } else {

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 ${COMMON_CMAKE_DEFS} ${CMAKE_DEFS}" CMAKE_DEFS="-DLLAMA_CUBLAS=on -DLLAMA_CUDA_FORCE_MMQ=on ${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