Request and model concurrency

This change adds support for multiple concurrent requests, as well as
loading multiple models by spawning multiple runners. The default
settings are currently set at 1 concurrent request per model and only 1
loaded model at a time, but these can be adjusted by setting
OLLAMA_NUM_PARALLEL and OLLAMA_MAX_LOADED_MODELS.

llm/payload.go

@@ -9,6 +9,7 @@ import (
 	"log/slog"
 	"os"
 	"path/filepath"
+	"runtime"
 	"strings"
 
 	"golang.org/x/exp/slices"
@@ -138,6 +139,23 @@ func serversForGpu(info gpu.GpuInfo) []string {
 	return servers
 }
 
+// Return the optimal server for this CPU architecture
+func serverForCpu() string {
+	if runtime.GOOS == "darwin" && runtime.GOARCH == "arm64" {
+		return "metal"
+	}
+	variant := gpu.GetCPUVariant()
+	availableServers := availableServers()
+	if variant != "" {
+		for cmp := range availableServers {
+			if cmp == "cpu_"+variant {
+				return cmp
+			}
+		}
+	}
+	return "cpu"
+}
+
 // extract extracts the embedded files to the target directory
 func extractFiles(targetDir string, glob string) error {
 	files, err := fs.Glob(libEmbed, glob)