Merge pull request #5128 from zhewang1-intc/fix_levelzero_empty_symbol_detect

Fix levelzero empty symbol detect
This commit is contained in:
Daniel Hiltgen 2024-06-19 08:33:16 -07:00 committed by GitHub
commit 2abebb2cbe
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 27 additions and 36 deletions

View file

@ -57,8 +57,6 @@ var (
SchedSpread bool
// Set via OLLAMA_TMPDIR in the environment
TmpDir string
// Set via OLLAMA_INTEL_GPU in the environment
IntelGpu bool
// Set via CUDA_VISIBLE_DEVICES in the environment
CudaVisibleDevices string
@ -103,7 +101,6 @@ func AsMap() map[string]EnvVar {
ret["ROCR_VISIBLE_DEVICES"] = EnvVar{"ROCR_VISIBLE_DEVICES", RocrVisibleDevices, "Set which AMD devices are visible"}
ret["GPU_DEVICE_ORDINAL"] = EnvVar{"GPU_DEVICE_ORDINAL", GpuDeviceOrdinal, "Set which AMD devices are visible"}
ret["HSA_OVERRIDE_GFX_VERSION"] = EnvVar{"HSA_OVERRIDE_GFX_VERSION", HsaOverrideGfxVersion, "Override the gfx used for all detected AMD GPUs"}
ret["OLLAMA_INTEL_GPU"] = EnvVar{"OLLAMA_INTEL_GPU", IntelGpu, "Enable experimental Intel GPU detection"}
}
return ret
}
@ -279,10 +276,6 @@ func LoadConfig() {
slog.Error("invalid setting", "OLLAMA_HOST", Host, "error", err, "using default port", Host.Port)
}
if set, err := strconv.ParseBool(clean("OLLAMA_INTEL_GPU")); err == nil {
IntelGpu = set
}
CudaVisibleDevices = clean("CUDA_VISIBLE_DEVICES")
HipVisibleDevices = clean("HIP_VISIBLE_DEVICES")
RocrVisibleDevices = clean("ROCR_VISIBLE_DEVICES")

View file

@ -280,35 +280,33 @@ func GetGPUInfo() GpuInfoList {
}
// Intel
if envconfig.IntelGpu {
oHandles = initOneAPIHandles()
for d := range oHandles.oneapi.num_drivers {
if oHandles.oneapi == nil {
// shouldn't happen
slog.Warn("nil oneapi handle with driver count", "count", int(oHandles.oneapi.num_drivers))
continue
}
devCount := C.oneapi_get_device_count(*oHandles.oneapi, C.int(d))
for i := range devCount {
gpuInfo := OneapiGPUInfo{
GpuInfo: GpuInfo{
Library: "oneapi",
},
driverIndex: int(d),
gpuIndex: int(i),
}
// TODO - split bootstrapping from updating free memory
C.oneapi_check_vram(*oHandles.oneapi, C.int(d), i, &memInfo)
// TODO - convert this to MinimumMemory based on testing...
var totalFreeMem float64 = float64(memInfo.free) * 0.95 // work-around: leave some reserve vram for mkl lib used in ggml-sycl backend.
memInfo.free = C.uint64_t(totalFreeMem)
gpuInfo.TotalMemory = uint64(memInfo.total)
gpuInfo.FreeMemory = uint64(memInfo.free)
gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
// TODO dependency path?
oneapiGPUs = append(oneapiGPUs, gpuInfo)
oHandles = initOneAPIHandles()
for d := 0; oHandles.oneapi != nil && d < int(oHandles.oneapi.num_drivers); d++ {
if oHandles.oneapi == nil {
// shouldn't happen
slog.Warn("nil oneapi handle with driver count", "count", int(oHandles.oneapi.num_drivers))
continue
}
devCount := C.oneapi_get_device_count(*oHandles.oneapi, C.int(d))
for i := range devCount {
gpuInfo := OneapiGPUInfo{
GpuInfo: GpuInfo{
Library: "oneapi",
},
driverIndex: d,
gpuIndex: int(i),
}
// TODO - split bootstrapping from updating free memory
C.oneapi_check_vram(*oHandles.oneapi, C.int(d), i, &memInfo)
// TODO - convert this to MinimumMemory based on testing...
var totalFreeMem float64 = float64(memInfo.free) * 0.95 // work-around: leave some reserve vram for mkl lib used in ggml-sycl backend.
memInfo.free = C.uint64_t(totalFreeMem)
gpuInfo.TotalMemory = uint64(memInfo.total)
gpuInfo.FreeMemory = uint64(memInfo.free)
gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
// TODO dependency path?
oneapiGPUs = append(oneapiGPUs, gpuInfo)
}
}

View file

@ -50,7 +50,7 @@ void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp) {
LOG(resp->oh.verbose, "dlsym: %s\n", l[i].s);
*l[i].p = LOAD_SYMBOL(resp->oh.handle, l[i].s);
if (!l[i].p) {
if (!*(l[i].p)) {
resp->oh.handle = NULL;
char *msg = LOAD_ERR();
LOG(resp->oh.verbose, "dlerr: %s\n", msg);