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

Revert "Limit GPU lib search for now (#4777)"

Browse source 
This reverts commit 476fb8e892.
This commit is contained in:
Daniel Hiltgen 2024-06-03 08:31:48 -07:00
parent 4dc7fb9525
commit efac488675
1 changed files with 31 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

32
gpu/gpu.go
View file

@ -16,12 +16,13 @@ import (
"os" "os"
"path/filepath" "path/filepath"
"runtime" "runtime"
"strconv"
"strings" "strings"
"sync" "sync"
"unsafe" "unsafe"
"github.com/ollama/ollama/envconfig"
"github.com/ollama/ollama/format" "github.com/ollama/ollama/format"
"github.com/ollama/ollama/envconfig"
) )
type handles struct { type handles struct {
@ -104,6 +105,8 @@ func initGPUHandles() *handles {
var cudartMgmtPatterns []string var cudartMgmtPatterns []string
var nvcudaMgmtName string var nvcudaMgmtName string
var nvcudaMgmtPatterns []string var nvcudaMgmtPatterns []string
var oneapiMgmtName string
var oneapiMgmtPatterns []string
tmpDir, _ := PayloadsDir() tmpDir, _ := PayloadsDir()
switch runtime.GOOS { switch runtime.GOOS {
@ -115,6 +118,8 @@ func initGPUHandles() *handles {
// Aligned with driver, we can't carry as payloads // Aligned with driver, we can't carry as payloads
nvcudaMgmtName = "nvcuda.dll" nvcudaMgmtName = "nvcuda.dll"
nvcudaMgmtPatterns = NvcudaWindowsGlobs nvcudaMgmtPatterns = NvcudaWindowsGlobs
oneapiMgmtName = "ze_intel_gpu64.dll"
oneapiMgmtPatterns = OneapiWindowsGlobs
case "linux": case "linux":
cudartMgmtName = "libcudart.so*" cudartMgmtName = "libcudart.so*"
if tmpDir != "" { if tmpDir != "" {
@ -125,6 +130,8 @@ func initGPUHandles() *handles {
// Aligned with driver, we can't carry as payloads // Aligned with driver, we can't carry as payloads
nvcudaMgmtName = "libcuda.so*" nvcudaMgmtName = "libcuda.so*"
nvcudaMgmtPatterns = NvcudaLinuxGlobs nvcudaMgmtPatterns = NvcudaLinuxGlobs
oneapiMgmtName = "libze_intel_gpu.so"
oneapiMgmtPatterns = OneapiLinuxGlobs
default: default:
return gpuHandles 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 return gpuHandles
} }
@ -227,6 +245,18 @@ func GetGPUInfo() GpuInfoList {
// TODO potentially sort on our own algorithm instead of what the underlying GPU library does... // TODO potentially sort on our own algorithm instead of what the underlying GPU library does...
resp = append(resp, gpuInfo) 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 // Then AMD