Get up and running with Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.
Revert "Limit GPU lib search for now (#4777)"
Daniel Hiltgenefac488675bb
diff --git a/gpu/gpu.go b/gpu/gpu.goindex a55903c51..7de99824c 100644--- a/gpu/gpu.go+++ b/gpu/gpu.go@@ -16,12 +16,13 @@ import ("os""path/filepath""runtime"+ "strconv""strings""sync""unsafe"- "github.com/ollama/ollama/envconfig""github.com/ollama/ollama/format"+ "github.com/ollama/ollama/envconfig")type handles struct {@@ -104,6 +105,8 @@ func initGPUHandles() *handles {var cudartMgmtPatterns []stringvar nvcudaMgmtName stringvar nvcudaMgmtPatterns []string+ var oneapiMgmtName string+ var oneapiMgmtPatterns []stringtmpDir, _ := PayloadsDir()switch runtime.GOOS {@@ -115,6 +118,8 @@ func initGPUHandles() *handles {// Aligned with driver, we can't carry as payloadsnvcudaMgmtName = "nvcuda.dll"nvcudaMgmtPatterns = NvcudaWindowsGlobs+ oneapiMgmtName = "ze_intel_gpu64.dll"+ oneapiMgmtPatterns = OneapiWindowsGlobscase "linux":cudartMgmtName = "libcudart.so*"if tmpDir != "" {@@ -125,6 +130,8 @@ func initGPUHandles() *handles {// Aligned with driver, we can't carry as payloadsnvcudaMgmtName = "libcuda.so*"nvcudaMgmtPatterns = NvcudaLinuxGlobs+ oneapiMgmtName = "libze_intel_gpu.so"+ oneapiMgmtPatterns = OneapiLinuxGlobsdefault:return gpuHandles}@@ -152,6 +159,17 @@ func initGPUHandles() *handles {}}+ oneapiLibPaths := FindGPULibs(oneapiMgmtName, oneapiMgmtPatterns)+ if len(oneapiLibPaths) > 0 {+ deviceCount, oneapi, libPath := LoadOneapiMgmt(oneapiLibPaths)+ if oneapi != nil {+ slog.Debug("detected Intel GPUs", "library", libPath, "count", deviceCount)+ gpuHandles.oneapi = oneapi+ gpuHandles.deviceCount = deviceCount+ return gpuHandles+ }+ }+return gpuHandles}@@ -227,6 +245,18 @@ func GetGPUInfo() GpuInfoList {// TODO potentially sort on our own algorithm instead of what the underlying GPU library does...resp = append(resp, gpuInfo)}+ if gpuHandles.oneapi != nil {+ gpuInfo := GpuInfo{+ Library: "oneapi",+ }+ C.oneapi_check_vram(*gpuHandles.oneapi, &memInfo)+ 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 = strconv.Itoa(i)+ resp = append(resp, gpuInfo)+ }}// Then AMD