feat(llama.cpp/clip): inject gpu options if we detect GPUs (#5243)

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

core/config/guesser.go

@@ -4,6 +4,7 @@ import (
 	"os"
 	"path/filepath"
 
+	"github.com/mudler/LocalAI/pkg/xsysinfo"
 	"github.com/rs/zerolog/log"
 	gguf "github.com/thxcode/gguf-parser-go"
 )
@@ -35,4 +36,10 @@ func guessDefaultsFromFile(cfg *BackendConfig, modelPath string, defaultCtx int)
 		}
 		cfg.ContextSize = &defaultCtx
 	}
+
+	if cfg.Options == nil {
+		if xsysinfo.HasGPU("nvidia") || xsysinfo.HasGPU("amd") {
+			cfg.Options = []string{"gpu"}
+		}
+	}
 }

pkg/model/initializers.go

@@ -181,10 +181,6 @@ func orderBackends(backends map[string][]string) ([]string, error) {
 // selectGRPCProcessByHostCapabilities selects the GRPC process to start based on system capabilities
 // Note: this is now relevant only for llama.cpp
 func selectGRPCProcessByHostCapabilities(backend, assetDir string, f16 bool) string {
-	foundCUDA := false
-	foundAMDGPU := false
-	foundIntelGPU := false
-	var grpcProcess string
 
 	// Select backend now just for llama.cpp
 	if backend != LLamaCPP {
@@ -198,48 +194,24 @@ func selectGRPCProcessByHostCapabilities(backend, assetDir string, f16 bool) str
 	}
 
 	// Check for GPU-binaries that are shipped with single binary releases
-	gpus, err := xsysinfo.GPUs()
+	gpuBinaries := map[string]string{
-	if err == nil {
+		"nvidia": LLamaCPPCUDA,
-		for _, gpu := range gpus {
+		"amd":    LLamaCPPHipblas,
-			if strings.Contains(gpu.String(), "nvidia") {
+		"intel":  LLamaCPPSycl16,
-				p := backendPath(assetDir, LLamaCPPCUDA)
-				if _, err := os.Stat(p); err == nil {
-					log.Info().Msgf("[%s] attempting to load with CUDA variant", backend)
-					grpcProcess = p
-					foundCUDA = true
-				} else {
-					log.Debug().Msgf("Nvidia GPU device found, no embedded CUDA variant found. You can ignore this message if you are using container with CUDA support")
-				}
-			}
-			if strings.Contains(gpu.String(), "amd") {
-				p := backendPath(assetDir, LLamaCPPHipblas)
-				if _, err := os.Stat(p); err == nil {
-					log.Info().Msgf("[%s] attempting to load with HIPBLAS variant", backend)
-					grpcProcess = p
-					foundAMDGPU = true
-				} else {
-					log.Debug().Msgf("AMD GPU device found, no embedded HIPBLAS variant found. You can ignore this message if you are using container with HIPBLAS support")
-				}
-			}
-			if strings.Contains(gpu.String(), "intel") {
-				backend := LLamaCPPSycl16
-				if !f16 {
-					backend = LLamaCPPSycl32
-				}
-				p := backendPath(assetDir, backend)
-				if _, err := os.Stat(p); err == nil {
-					log.Info().Msgf("[%s] attempting to load with Intel variant", backend)
-					grpcProcess = p
-					foundIntelGPU = true
-				} else {
-					log.Debug().Msgf("Intel GPU device found, no embedded SYCL variant found. You can ignore this message if you are using container with SYCL support")
-				}
-			}
-		}
 	}
 
-	if foundCUDA || foundAMDGPU || foundIntelGPU {
+	if !f16 {
-		return grpcProcess
+		gpuBinaries["intel"] = LLamaCPPSycl32
+	}
+
+	for vendor, binary := range gpuBinaries {
+		if xsysinfo.HasGPU(vendor) {
+			p := backendPath(assetDir, binary)
+			if _, err := os.Stat(p); err == nil {
+				log.Info().Msgf("[%s] attempting to load with %s variant (vendor: %s)", backend, binary, vendor)
+				return p
+			}
+		}
 	}
 
 	// No GPU found or no specific binaries found, try to load the CPU variant(s)

pkg/xsysinfo/gpu.go

@@ -1,6 +1,8 @@
 package xsysinfo
 
 import (
+	"strings"
+
 	"github.com/jaypipes/ghw"
 	"github.com/jaypipes/ghw/pkg/gpu"
 )
@@ -13,3 +15,19 @@ func GPUs() ([]*gpu.GraphicsCard, error) {
 
 	return gpu.GraphicsCards, nil
 }
+
+func HasGPU(vendor string) bool {
+	gpus, err := GPUs()
+	if err != nil {
+		return false
+	}
+	if vendor == "" {
+		return len(gpus) > 0
+	}
+	for _, gpu := range gpus {
+		if strings.Contains(gpu.String(), vendor) {
+			return true
+		}
+	}
+	return false
+}