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

Revert "Revert "gpu: add env var for detecting Intel oneapi gpus (#5076)""

Browse source 
This reverts commit 755b4e4fc2.
This commit is contained in:
Daniel Hiltgen 2024-06-19 08:57:41 -07:00
parent 2abebb2cbe
commit d34d88e417
2 changed files with 35 additions and 26 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

7
envconfig/config.go
View file

@ -57,6 +57,8 @@ var (
SchedSpread bool SchedSpread bool
// Set via OLLAMA_TMPDIR in the environment // Set via OLLAMA_TMPDIR in the environment
TmpDir string TmpDir string
// Set via OLLAMA_INTEL_GPU in the environment
IntelGpu bool
// Set via CUDA_VISIBLE_DEVICES in the environment // Set via CUDA_VISIBLE_DEVICES in the environment
CudaVisibleDevices string CudaVisibleDevices string
@ -101,6 +103,7 @@ func AsMap() map[string]EnvVar {
ret["ROCR_VISIBLE_DEVICES"] = EnvVar{"ROCR_VISIBLE_DEVICES", RocrVisibleDevices, "Set which AMD devices are visible"} 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["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["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 return ret
} }
@ -276,6 +279,10 @@ func LoadConfig() {
slog.Error("invalid setting", "OLLAMA_HOST", Host, "error", err, "using default port", Host.Port) 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") CudaVisibleDevices = clean("CUDA_VISIBLE_DEVICES")
HipVisibleDevices = clean("HIP_VISIBLE_DEVICES") HipVisibleDevices = clean("HIP_VISIBLE_DEVICES")
RocrVisibleDevices = clean("ROCR_VISIBLE_DEVICES") RocrVisibleDevices = clean("ROCR_VISIBLE_DEVICES")

54
gpu/gpu.go
View file

@ -280,33 +280,35 @@ func GetGPUInfo() GpuInfoList {
} }
// Intel // Intel
oHandles = initOneAPIHandles() if envconfig.IntelGpu {
for d := 0; oHandles.oneapi != nil && d < int(oHandles.oneapi.num_drivers); d++ { oHandles = initOneAPIHandles()
if oHandles.oneapi == nil { for d := range oHandles.oneapi.num_drivers {
// shouldn't happen if oHandles.oneapi == nil {
slog.Warn("nil oneapi handle with driver count", "count", int(oHandles.oneapi.num_drivers)) // shouldn't happen
continue 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 { devCount := C.oneapi_get_device_count(*oHandles.oneapi, C.int(d))
gpuInfo := OneapiGPUInfo{ for i := range devCount {
GpuInfo: GpuInfo{ gpuInfo := OneapiGPUInfo{
Library: "oneapi", GpuInfo: GpuInfo{
}, Library: "oneapi",
driverIndex: d, },
gpuIndex: int(i), 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)
} }
// 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)
} }
} }