mirror of
https://github.com/dogkeeper886/ollama37.git
synced 2025-12-12 00:37:04 +00:00
Track GPU discovery failure information (#5820)
* Expose GPU discovery failure information * Remove exposed API for now
This commit is contained in:
145
gpu/gpu.go
145
gpu/gpu.go
@@ -54,6 +54,13 @@ var (
|
||||
nvmlLibPath string
|
||||
rocmGPUs []RocmGPUInfo
|
||||
oneapiGPUs []OneapiGPUInfo
|
||||
|
||||
// If any discovered GPUs are incompatible, report why
|
||||
unsupportedGPUs []UnsupportedGPUInfo
|
||||
|
||||
// Keep track of errors during bootstrapping so that if GPUs are missing
|
||||
// they expected to be present this may explain why
|
||||
bootstrapErrors []error
|
||||
)
|
||||
|
||||
// With our current CUDA compile flags, older than 5.0 will not work properly
|
||||
@@ -70,16 +77,17 @@ func initCudaHandles() *cudaHandles {
|
||||
|
||||
cHandles := &cudaHandles{}
|
||||
// Short Circuit if we already know which library to use
|
||||
// ignore bootstrap errors in this case since we already recorded them
|
||||
if nvmlLibPath != "" {
|
||||
cHandles.nvml, _ = LoadNVMLMgmt([]string{nvmlLibPath})
|
||||
cHandles.nvml, _, _ = loadNVMLMgmt([]string{nvmlLibPath})
|
||||
return cHandles
|
||||
}
|
||||
if nvcudaLibPath != "" {
|
||||
cHandles.deviceCount, cHandles.nvcuda, _ = LoadNVCUDAMgmt([]string{nvcudaLibPath})
|
||||
cHandles.deviceCount, cHandles.nvcuda, _, _ = loadNVCUDAMgmt([]string{nvcudaLibPath})
|
||||
return cHandles
|
||||
}
|
||||
if cudartLibPath != "" {
|
||||
cHandles.deviceCount, cHandles.cudart, _ = LoadCUDARTMgmt([]string{cudartLibPath})
|
||||
cHandles.deviceCount, cHandles.cudart, _, _ = loadCUDARTMgmt([]string{cudartLibPath})
|
||||
return cHandles
|
||||
}
|
||||
|
||||
@@ -102,18 +110,21 @@ func initCudaHandles() *cudaHandles {
|
||||
if len(NvmlGlobs) > 0 {
|
||||
nvmlLibPaths := FindGPULibs(NvmlMgmtName, NvmlGlobs)
|
||||
if len(nvmlLibPaths) > 0 {
|
||||
nvml, libPath := LoadNVMLMgmt(nvmlLibPaths)
|
||||
nvml, libPath, err := loadNVMLMgmt(nvmlLibPaths)
|
||||
if nvml != nil {
|
||||
slog.Debug("nvidia-ml loaded", "library", libPath)
|
||||
cHandles.nvml = nvml
|
||||
nvmlLibPath = libPath
|
||||
}
|
||||
if err != nil {
|
||||
bootstrapErrors = append(bootstrapErrors, err)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
nvcudaLibPaths := FindGPULibs(NvcudaMgmtName, nvcudaMgmtPatterns)
|
||||
if len(nvcudaLibPaths) > 0 {
|
||||
deviceCount, nvcuda, libPath := LoadNVCUDAMgmt(nvcudaLibPaths)
|
||||
deviceCount, nvcuda, libPath, err := loadNVCUDAMgmt(nvcudaLibPaths)
|
||||
if nvcuda != nil {
|
||||
slog.Debug("detected GPUs", "count", deviceCount, "library", libPath)
|
||||
cHandles.nvcuda = nvcuda
|
||||
@@ -121,11 +132,14 @@ func initCudaHandles() *cudaHandles {
|
||||
nvcudaLibPath = libPath
|
||||
return cHandles
|
||||
}
|
||||
if err != nil {
|
||||
bootstrapErrors = append(bootstrapErrors, err)
|
||||
}
|
||||
}
|
||||
|
||||
cudartLibPaths := FindGPULibs(CudartMgmtName, cudartMgmtPatterns)
|
||||
if len(cudartLibPaths) > 0 {
|
||||
deviceCount, cudart, libPath := LoadCUDARTMgmt(cudartLibPaths)
|
||||
deviceCount, cudart, libPath, err := loadCUDARTMgmt(cudartLibPaths)
|
||||
if cudart != nil {
|
||||
slog.Debug("detected GPUs", "library", libPath, "count", deviceCount)
|
||||
cHandles.cudart = cudart
|
||||
@@ -133,6 +147,9 @@ func initCudaHandles() *cudaHandles {
|
||||
cudartLibPath = libPath
|
||||
return cHandles
|
||||
}
|
||||
if err != nil {
|
||||
bootstrapErrors = append(bootstrapErrors, err)
|
||||
}
|
||||
}
|
||||
|
||||
return cHandles
|
||||
@@ -143,14 +160,19 @@ func initOneAPIHandles() *oneapiHandles {
|
||||
oHandles := &oneapiHandles{}
|
||||
|
||||
// Short Circuit if we already know which library to use
|
||||
// ignore bootstrap errors in this case since we already recorded them
|
||||
if oneapiLibPath != "" {
|
||||
oHandles.deviceCount, oHandles.oneapi, _ = LoadOneapiMgmt([]string{oneapiLibPath})
|
||||
oHandles.deviceCount, oHandles.oneapi, _, _ = loadOneapiMgmt([]string{oneapiLibPath})
|
||||
return oHandles
|
||||
}
|
||||
|
||||
oneapiLibPaths := FindGPULibs(OneapiMgmtName, OneapiGlobs)
|
||||
if len(oneapiLibPaths) > 0 {
|
||||
oHandles.deviceCount, oHandles.oneapi, oneapiLibPath = LoadOneapiMgmt(oneapiLibPaths)
|
||||
var err error
|
||||
oHandles.deviceCount, oHandles.oneapi, oneapiLibPath, err = loadOneapiMgmt(oneapiLibPaths)
|
||||
if err != nil {
|
||||
bootstrapErrors = append(bootstrapErrors, err)
|
||||
}
|
||||
}
|
||||
|
||||
return oHandles
|
||||
@@ -197,6 +219,7 @@ func GetGPUInfo() GpuInfoList {
|
||||
|
||||
if !bootstrapped {
|
||||
slog.Info("looking for compatible GPUs")
|
||||
bootstrapErrors = []error{}
|
||||
needRefresh = false
|
||||
cpuCapability = GetCPUCapability()
|
||||
var memInfo C.mem_info_t
|
||||
@@ -221,7 +244,9 @@ func GetGPUInfo() GpuInfoList {
|
||||
|
||||
// Fallback to CPU mode if we're lacking required vector extensions on x86
|
||||
if cpuCapability < GPURunnerCPUCapability && runtime.GOARCH == "amd64" {
|
||||
slog.Warn("CPU does not have minimum vector extensions, GPU inference disabled", "required", GPURunnerCPUCapability, "detected", cpuCapability)
|
||||
err := fmt.Errorf("CPU does not have minimum vector extensions, GPU inference disabled. Required:%s Detected:%s", GPURunnerCPUCapability, cpuCapability)
|
||||
slog.Warn(err.Error())
|
||||
bootstrapErrors = append(bootstrapErrors, err)
|
||||
bootstrapped = true
|
||||
// No need to do any GPU discovery, since we can't run on them
|
||||
return GpuInfoList{cpus[0].GpuInfo}
|
||||
@@ -253,10 +278,6 @@ func GetGPUInfo() GpuInfoList {
|
||||
C.free(unsafe.Pointer(memInfo.err))
|
||||
continue
|
||||
}
|
||||
if memInfo.major < CudaComputeMin[0] || (memInfo.major == CudaComputeMin[0] && memInfo.minor < CudaComputeMin[1]) {
|
||||
slog.Info(fmt.Sprintf("[%d] CUDA GPU is too old. Compute Capability detected: %d.%d", i, memInfo.major, memInfo.minor))
|
||||
continue
|
||||
}
|
||||
gpuInfo.TotalMemory = uint64(memInfo.total)
|
||||
gpuInfo.FreeMemory = uint64(memInfo.free)
|
||||
gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
|
||||
@@ -279,6 +300,15 @@ func GetGPUInfo() GpuInfoList {
|
||||
gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
|
||||
gpuInfo.Variant = variant
|
||||
|
||||
if memInfo.major < CudaComputeMin[0] || (memInfo.major == CudaComputeMin[0] && memInfo.minor < CudaComputeMin[1]) {
|
||||
unsupportedGPUs = append(unsupportedGPUs,
|
||||
UnsupportedGPUInfo{
|
||||
GpuInfo: gpuInfo.GpuInfo,
|
||||
})
|
||||
slog.Info(fmt.Sprintf("[%d] CUDA GPU is too old. Compute Capability detected: %d.%d", i, memInfo.major, memInfo.minor))
|
||||
continue
|
||||
}
|
||||
|
||||
// query the management library as well so we can record any skew between the two
|
||||
// which represents overhead on the GPU we must set aside on subsequent updates
|
||||
if cHandles.nvml != nil {
|
||||
@@ -341,7 +371,10 @@ func GetGPUInfo() GpuInfoList {
|
||||
}
|
||||
}
|
||||
|
||||
rocmGPUs = AMDGetGPUInfo()
|
||||
rocmGPUs, err = AMDGetGPUInfo()
|
||||
if err != nil {
|
||||
bootstrapErrors = append(bootstrapErrors, err)
|
||||
}
|
||||
bootstrapped = true
|
||||
if len(cudaGPUs) == 0 && len(rocmGPUs) == 0 && len(oneapiGPUs) == 0 {
|
||||
slog.Info("no compatible GPUs were discovered")
|
||||
@@ -526,92 +559,114 @@ func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
|
||||
return gpuLibPaths
|
||||
}
|
||||
|
||||
func LoadCUDARTMgmt(cudartLibPaths []string) (int, *C.cudart_handle_t, string) {
|
||||
// Bootstrap the runtime library
|
||||
// Returns: num devices, handle, libPath, error
|
||||
func loadCUDARTMgmt(cudartLibPaths []string) (int, *C.cudart_handle_t, string, error) {
|
||||
var resp C.cudart_init_resp_t
|
||||
resp.ch.verbose = getVerboseState()
|
||||
var err error
|
||||
for _, libPath := range cudartLibPaths {
|
||||
lib := C.CString(libPath)
|
||||
defer C.free(unsafe.Pointer(lib))
|
||||
C.cudart_init(lib, &resp)
|
||||
if resp.err != nil {
|
||||
slog.Debug("Unable to load cudart", "library", libPath, "error", C.GoString(resp.err))
|
||||
err = fmt.Errorf("Unable to load cudart library %s: %s", libPath, C.GoString(resp.err))
|
||||
slog.Debug(err.Error())
|
||||
C.free(unsafe.Pointer(resp.err))
|
||||
} else {
|
||||
return int(resp.num_devices), &resp.ch, libPath
|
||||
err = nil
|
||||
return int(resp.num_devices), &resp.ch, libPath, err
|
||||
}
|
||||
}
|
||||
return 0, nil, ""
|
||||
return 0, nil, "", err
|
||||
}
|
||||
|
||||
func LoadNVCUDAMgmt(nvcudaLibPaths []string) (int, *C.nvcuda_handle_t, string) {
|
||||
// Bootstrap the driver library
|
||||
// Returns: num devices, handle, libPath, error
|
||||
func loadNVCUDAMgmt(nvcudaLibPaths []string) (int, *C.nvcuda_handle_t, string, error) {
|
||||
var resp C.nvcuda_init_resp_t
|
||||
resp.ch.verbose = getVerboseState()
|
||||
var err error
|
||||
for _, libPath := range nvcudaLibPaths {
|
||||
lib := C.CString(libPath)
|
||||
defer C.free(unsafe.Pointer(lib))
|
||||
C.nvcuda_init(lib, &resp)
|
||||
if resp.err != nil {
|
||||
// Decide what log level based on the type of error message to help users understand why
|
||||
msg := C.GoString(resp.err)
|
||||
switch resp.cudaErr {
|
||||
case C.CUDA_ERROR_INSUFFICIENT_DRIVER, C.CUDA_ERROR_SYSTEM_DRIVER_MISMATCH:
|
||||
slog.Warn("version mismatch between driver and cuda driver library - reboot or upgrade may be required", "library", libPath, "error", msg)
|
||||
err = fmt.Errorf("version mismatch between driver and cuda driver library - reboot or upgrade may be required: library %s", libPath)
|
||||
slog.Warn(err.Error())
|
||||
case C.CUDA_ERROR_NO_DEVICE:
|
||||
slog.Info("no nvidia devices detected", "library", libPath)
|
||||
err = fmt.Errorf("no nvidia devices detected by library %s", libPath)
|
||||
slog.Info(err.Error())
|
||||
case C.CUDA_ERROR_UNKNOWN:
|
||||
slog.Warn("unknown error initializing cuda driver library", "library", libPath, "error", msg)
|
||||
slog.Warn("see https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for more information")
|
||||
err = fmt.Errorf("unknown error initializing cuda driver library %s: %s. see https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for more information", libPath, C.GoString(resp.err))
|
||||
slog.Warn(err.Error())
|
||||
default:
|
||||
msg := C.GoString(resp.err)
|
||||
if strings.Contains(msg, "wrong ELF class") {
|
||||
slog.Debug("skipping 32bit library", "library", libPath)
|
||||
} else {
|
||||
slog.Info("unable to load cuda driver library", "library", libPath, "error", msg)
|
||||
err = fmt.Errorf("Unable to load cudart library %s: %s", libPath, C.GoString(resp.err))
|
||||
slog.Info(err.Error())
|
||||
}
|
||||
}
|
||||
C.free(unsafe.Pointer(resp.err))
|
||||
} else {
|
||||
return int(resp.num_devices), &resp.ch, libPath
|
||||
err = nil
|
||||
return int(resp.num_devices), &resp.ch, libPath, err
|
||||
}
|
||||
}
|
||||
return 0, nil, ""
|
||||
return 0, nil, "", err
|
||||
}
|
||||
|
||||
func LoadNVMLMgmt(nvmlLibPaths []string) (*C.nvml_handle_t, string) {
|
||||
// Bootstrap the management library
|
||||
// Returns: handle, libPath, error
|
||||
func loadNVMLMgmt(nvmlLibPaths []string) (*C.nvml_handle_t, string, error) {
|
||||
var resp C.nvml_init_resp_t
|
||||
resp.ch.verbose = getVerboseState()
|
||||
var err error
|
||||
for _, libPath := range nvmlLibPaths {
|
||||
lib := C.CString(libPath)
|
||||
defer C.free(unsafe.Pointer(lib))
|
||||
C.nvml_init(lib, &resp)
|
||||
if resp.err != nil {
|
||||
slog.Info(fmt.Sprintf("Unable to load NVML management library %s: %s", libPath, C.GoString(resp.err)))
|
||||
err = fmt.Errorf("Unable to load NVML management library %s: %s", libPath, C.GoString(resp.err))
|
||||
slog.Info(err.Error())
|
||||
C.free(unsafe.Pointer(resp.err))
|
||||
} else {
|
||||
return &resp.ch, libPath
|
||||
err = nil
|
||||
return &resp.ch, libPath, err
|
||||
}
|
||||
}
|
||||
return nil, ""
|
||||
return nil, "", err
|
||||
}
|
||||
|
||||
func LoadOneapiMgmt(oneapiLibPaths []string) (int, *C.oneapi_handle_t, string) {
|
||||
// bootstrap the Intel GPU library
|
||||
// Returns: num devices, handle, libPath, error
|
||||
func loadOneapiMgmt(oneapiLibPaths []string) (int, *C.oneapi_handle_t, string, error) {
|
||||
var resp C.oneapi_init_resp_t
|
||||
num_devices := 0
|
||||
resp.oh.verbose = getVerboseState()
|
||||
var err error
|
||||
for _, libPath := range oneapiLibPaths {
|
||||
lib := C.CString(libPath)
|
||||
defer C.free(unsafe.Pointer(lib))
|
||||
C.oneapi_init(lib, &resp)
|
||||
if resp.err != nil {
|
||||
slog.Debug("Unable to load oneAPI management library", "library", libPath, "error", C.GoString(resp.err))
|
||||
err = fmt.Errorf("Unable to load oneAPI management library %s: %s", libPath, C.GoString(resp.err))
|
||||
slog.Debug(err.Error())
|
||||
C.free(unsafe.Pointer(resp.err))
|
||||
} else {
|
||||
err = nil
|
||||
for i := range resp.oh.num_drivers {
|
||||
num_devices += int(C.oneapi_get_device_count(resp.oh, C.int(i)))
|
||||
}
|
||||
return num_devices, &resp.oh, libPath
|
||||
return num_devices, &resp.oh, libPath, err
|
||||
}
|
||||
}
|
||||
return 0, nil, ""
|
||||
return 0, nil, "", err
|
||||
}
|
||||
|
||||
func getVerboseState() C.uint16_t {
|
||||
@@ -669,3 +724,23 @@ func LibraryDir() string {
|
||||
slog.Warn("unable to locate gpu dependency libraries")
|
||||
return ""
|
||||
}
|
||||
|
||||
func GetSystemInfo() SystemInfo {
|
||||
gpus := GetGPUInfo()
|
||||
gpuMutex.Lock()
|
||||
defer gpuMutex.Unlock()
|
||||
discoveryErrors := []string{}
|
||||
for _, err := range bootstrapErrors {
|
||||
discoveryErrors = append(discoveryErrors, err.Error())
|
||||
}
|
||||
if len(gpus) == 1 && gpus[0].Library == "cpu" {
|
||||
gpus = []GpuInfo{}
|
||||
}
|
||||
|
||||
return SystemInfo{
|
||||
System: cpus[0],
|
||||
GPUs: gpus,
|
||||
UnsupportedGPUs: unsupportedGPUs,
|
||||
DiscoveryErrors: discoveryErrors,
|
||||
}
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user