Skip to content

Get up and running with Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.

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

Daniel Hiltgend34d88e41744
diff --git a/envconfig/config.go b/envconfig/config.go
index bcf2e18ae..e86f72e6a 100644
--- a/envconfig/config.go
+++ b/envconfig/config.go
@@ -57,6 +57,8 @@ 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
@@ -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["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
}
@@ -276,6 +279,10 @@ 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")
diff --git a/gpu/gpu.go b/gpu/gpu.go
index 56a4dbfac..ce0a10496 100644
--- a/gpu/gpu.go
+++ b/gpu/gpu.go
@@ -280,33 +280,35 @@ func GetGPUInfo() GpuInfoList {
}
// Intel
- 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),
+ 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)
}
- // 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)
}
}