Always dynamically load the llm server library

This switches darwin to dynamic loading, and refactors the code now that no
static linking of the library is used on any platform

gpu/cpu_common.go

@@ -0,0 +1,21 @@
+package gpu
+
+import (
+	"log"
+
+	"golang.org/x/sys/cpu"
+)
+
+func GetCPUVariant() string {
+	if cpu.X86.HasAVX2 {
+		log.Printf("CPU has AVX2")
+		return "avx2"
+	}
+	if cpu.X86.HasAVX {
+		log.Printf("CPU has AVX")
+		return "avx"
+	}
+	log.Printf("CPU does not have vector extensions")
+	// else LCD
+	return ""
+}

gpu/gpu_darwin.go

@@ -32,8 +32,15 @@ func CheckVRAM() (int64, error) {
 
 func GetGPUInfo() GpuInfo {
 	mem, _ := getCPUMem()
+	if runtime.GOARCH == "amd64" {
+		return GpuInfo{
+			Library: "default",
+			Variant: GetCPUVariant(),
+			memInfo: mem,
+		}
+	}
 	return GpuInfo{
-		Library: "default",
+		Library: "metal",
 		memInfo: mem,
 	}
 }
@@ -45,12 +52,3 @@ func getCPUMem() (memInfo, error) {
 		DeviceCount: 0,
 	}, nil
 }
-
-func nativeInit() error {
-	return nil
-}
-
-func GetCPUVariant() string {
-	// We don't yet have CPU based builds for Darwin...
-	return ""
-}

gpu/gpu_test.go

@@ -9,7 +9,7 @@ import (
 
 func TestBasicGetGPUInfo(t *testing.T) {
 	info := GetGPUInfo()
-	assert.Contains(t, "cuda rocm cpu default", info.Library)
+	assert.Contains(t, "cuda rocm cpu metal", info.Library)
 
 	switch runtime.GOOS {
 	case "darwin":