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

llm/shim_ext_server_linux.go

-package llm
-
-import (
-	"embed"
-	"log"
-	"os"
-	"strings"
-)
-
-//go:embed llama.cpp/build/*/*/lib/*.so
-var libEmbed embed.FS
-
-func updatePath(dir string) {
-	pathComponents := strings.Split(os.Getenv("LD_LIBRARY_PATH"), ":")
-	for _, comp := range pathComponents {
-		if comp == dir {
-			return
-		}
-	}
-	newPath := strings.Join(append([]string{dir}, pathComponents...), ":")
-	log.Printf("Updating LD_LIBRARY_PATH to %s", newPath)
-	os.Setenv("LD_LIBRARY_PATH", newPath)
-}

llm/shim_ext_server_windows.go

-package llm
-
-import (
-	"embed"
-	"log"
-	"os"
-	"path/filepath"
-	"strings"
-)
-
-//go:embed llama.cpp/build/windows/*/lib/*.dll
-var libEmbed embed.FS
-
-func updatePath(dir string) {
-	tmpDir := filepath.Dir(dir)
-	pathComponents := strings.Split(os.Getenv("PATH"), ";")
-	i := 0
-	for _, comp := range pathComponents {
-		if strings.EqualFold(comp, dir) {
-			return
-		}
-		// Remove any other prior paths to our temp dir
-		if !strings.HasPrefix(strings.ToLower(comp), strings.ToLower(tmpDir)) {
-			pathComponents[i] = comp
-			i++
-		}
-	}
-	newPath := strings.Join(append([]string{dir}, pathComponents...), ";")
-	log.Printf("Updating PATH to %s", newPath)
-	os.Setenv("PATH", newPath)
-}

llm/shim_test.go

-package llm
-
-import (
-	"testing"
-
-	"github.com/jmorganca/ollama/gpu"
-	"github.com/stretchr/testify/assert"
-)
-
-func TestGetShims(t *testing.T) {
-	availableShims = map[string]string{
-		"cpu": "X_cpu",
-	}
-	assert.Equal(t, false, rocmShimPresent())
-	res := getShims(gpu.GpuInfo{Library: "cpu"})
-	assert.Len(t, res, 1)
-	assert.Equal(t, availableShims["cpu"], res[0])
-
-	availableShims = map[string]string{
-		"rocm_v5": "X_rocm_v5",
-		"rocm_v6": "X_rocm_v6",
-		"cpu":     "X_cpu",
-	}
-	assert.Equal(t, true, rocmShimPresent())
-	res = getShims(gpu.GpuInfo{Library: "rocm"})
-	assert.Len(t, res, 3)
-	assert.Equal(t, availableShims["rocm_v5"], res[0])
-	assert.Equal(t, availableShims["rocm_v6"], res[1])
-	assert.Equal(t, availableShims["cpu"], res[2])
-
-	res = getShims(gpu.GpuInfo{Library: "rocm", Variant: "v6"})
-	assert.Len(t, res, 3)
-	assert.Equal(t, availableShims["rocm_v6"], res[0])
-	assert.Equal(t, availableShims["rocm_v5"], res[1])
-	assert.Equal(t, availableShims["cpu"], res[2])
-
-	res = getShims(gpu.GpuInfo{Library: "cuda"})
-	assert.Len(t, res, 1)
-	assert.Equal(t, availableShims["cpu"], res[0])
-
-	res = getShims(gpu.GpuInfo{Library: "default"})
-	assert.Len(t, res, 1)
-	assert.Equal(t, "default", res[0])
-
-	availableShims = map[string]string{
-		"rocm": "X_rocm_v5",
-		"cpu":  "X_cpu",
-	}
-	assert.Equal(t, true, rocmShimPresent())
-	res = getShims(gpu.GpuInfo{Library: "rocm", Variant: "v6"})
-	assert.Len(t, res, 2)
-	assert.Equal(t, availableShims["rocm"], res[0])
-	assert.Equal(t, availableShims["cpu"], res[1])
-}