Added support for proxying llama.cpp native API endpoints via /llama-cpp/{name}/

pkg/server/handlers.go

@@ -480,29 +480,15 @@ func (h *Handler) ProxyToInstance() http.HandlerFunc {
 
 		// Strip the "/api/v1/instances/<name>/proxy" prefix from the request URL
 		prefix := fmt.Sprintf("/api/v1/instances/%s/proxy", name)
-		proxyPath := r.URL.Path[len(prefix):]
-
-		// Ensure the proxy path starts with "/"
-		if !strings.HasPrefix(proxyPath, "/") {
-			proxyPath = "/" + proxyPath
-		}
+		r.URL.Path = strings.TrimPrefix(r.URL.Path, prefix)
 
 		// Update the last request time for the instance
 		inst.UpdateLastRequestTime()
 
-		// Modify the request to remove the proxy prefix
-		originalPath := r.URL.Path
-		r.URL.Path = proxyPath
-
 		// Set forwarded headers
 		r.Header.Set("X-Forwarded-Host", r.Header.Get("Host"))
 		r.Header.Set("X-Forwarded-Proto", "http")
 
-		// Restore original path for logging purposes
-		defer func() {
-			r.URL.Path = originalPath
-		}()
-
 		// Forward the request using the cached proxy
 		proxy.ServeHTTP(w, r)
 	}
@@ -585,12 +571,13 @@ func (h *Handler) OpenAIProxy() http.HandlerFunc {
 		// Route to the appropriate inst based on instance name
 		inst, err := h.InstanceManager.GetInstance(modelName)
 		if err != nil {
-			http.Error(w, "Failed to get instance: "+err.Error(), http.StatusInternalServerError)
+			http.Error(w, "Invalid instance: "+err.Error(), http.StatusBadRequest)
 			return
 		}
 
 		if !inst.IsRunning() {
-			allowOnDemand := inst.GetOptions() != nil && inst.GetOptions().OnDemandStart != nil && *inst.GetOptions().OnDemandStart
+			options := inst.GetOptions()
+			allowOnDemand := options != nil && options.OnDemandStart != nil && *options.OnDemandStart
 			if !allowOnDemand {
 				http.Error(w, "Instance is not running", http.StatusServiceUnavailable)
 				return
@@ -639,6 +626,84 @@ func (h *Handler) OpenAIProxy() http.HandlerFunc {
 	}
 }
 
+func (h *Handler) LlamaCppProxy(onDemandStart bool) http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+
+		// Get the instance name from the URL parameter
+		name := chi.URLParam(r, "name")
+		if name == "" {
+			http.Error(w, "Instance name cannot be empty", http.StatusBadRequest)
+			return
+		}
+
+		// Route to the appropriate inst based on instance name
+		inst, err := h.InstanceManager.GetInstance(name)
+		if err != nil {
+			http.Error(w, "Invalid instance: "+err.Error(), http.StatusBadRequest)
+			return
+		}
+
+		options := inst.GetOptions()
+		if options == nil {
+			http.Error(w, "Cannot obtain Instance's options", http.StatusInternalServerError)
+			return
+		}
+
+		if options.BackendType != backends.BackendTypeLlamaCpp {
+			http.Error(w, "Instance is not a llama.cpp server.", http.StatusBadRequest)
+			return
+		}
+
+		if !inst.IsRunning() {
+
+			if !(onDemandStart && options.OnDemandStart != nil && *options.OnDemandStart) {
+				http.Error(w, "Instance is not running", http.StatusServiceUnavailable)
+				return
+			}
+
+			if h.InstanceManager.IsMaxRunningInstancesReached() {
+				if h.cfg.Instances.EnableLRUEviction {
+					err := h.InstanceManager.EvictLRUInstance()
+					if err != nil {
+						http.Error(w, "Cannot start Instance, failed to evict instance "+err.Error(), http.StatusInternalServerError)
+						return
+					}
+				} else {
+					http.Error(w, "Cannot start Instance, maximum number of instances reached", http.StatusConflict)
+					return
+				}
+			}
+
+			// If on-demand start is enabled, start the instance
+			if _, err := h.InstanceManager.StartInstance(name); err != nil {
+				http.Error(w, "Failed to start instance: "+err.Error(), http.StatusInternalServerError)
+				return
+			}
+
+			// Wait for the instance to become healthy before proceeding
+			if err := inst.WaitForHealthy(h.cfg.Instances.OnDemandStartTimeout); err != nil { // 2 minutes timeout
+				http.Error(w, "Instance failed to become healthy: "+err.Error(), http.StatusServiceUnavailable)
+				return
+			}
+		}
+
+		proxy, err := inst.GetProxy()
+		if err != nil {
+			http.Error(w, "Failed to get proxy: "+err.Error(), http.StatusInternalServerError)
+			return
+		}
+
+		// Strip the "/llama-cpp/<name>" prefix from the request URL
+		prefix := fmt.Sprintf("/llama-cpp/%s", name)
+		r.URL.Path = strings.TrimPrefix(r.URL.Path, prefix)
+
+		// Update the last request time for the instance
+		inst.UpdateLastRequestTime()
+
+		proxy.ServeHTTP(w, r)
+	}
+}
+
 // ParseCommandRequest represents the request body for command parsing
 type ParseCommandRequest struct {
 	Command string `json:"command"`
@@ -719,21 +784,21 @@ func (h *Handler) ParseMlxCommand() http.HandlerFunc {
 			writeError(w, http.StatusBadRequest, "invalid_request", "Invalid JSON body")
 			return
 		}
-		
+
 		if strings.TrimSpace(req.Command) == "" {
 			writeError(w, http.StatusBadRequest, "invalid_command", "Command cannot be empty")
 			return
 		}
-		
+
 		mlxOptions, err := mlx.ParseMlxCommand(req.Command)
 		if err != nil {
 			writeError(w, http.StatusBadRequest, "parse_error", err.Error())
 			return
 		}
-		
+
 		// Currently only support mlx_lm backend type
 		backendType := backends.BackendTypeMlxLm
-		
+
 		options := &instance.CreateInstanceOptions{
 			BackendType:      backendType,
 			MlxServerOptions: mlxOptions,

pkg/server/routes.go

@@ -103,6 +103,51 @@ func SetupRouter(handler *Handler) *chi.Mux {
 
 	})
 
+	r.Route("/llama-cpp/{name}", func(r chi.Router) {
+
+		// Public Routes
+		// Allow llama-cpp server to serve its own WebUI if it is running.
+		// Don't auto start the server since it can be accessed without an API key
+		r.Get("/", handler.LlamaCppProxy(false))
+
+		// Private Routes
+		r.Group(func(r chi.Router) {
+
+			if authMiddleware != nil && handler.cfg.Auth.RequireInferenceAuth {
+				r.Use(authMiddleware.AuthMiddleware(KeyTypeInference))
+			}
+
+			// This handler auto start the server if it's not running
+			llamaCppHandler := handler.LlamaCppProxy(true)
+
+			// llama.cpp server specific proxy endpoints
+			r.Get("/props", llamaCppHandler)
+			// /slots endpoint is secured (see: https://github.com/ggml-org/llama.cpp/pull/15630)
+			r.Get("/slots", llamaCppHandler)
+			r.Post("/apply-template", llamaCppHandler)
+			r.Post("/completion", llamaCppHandler)
+			r.Post("/detokenize", llamaCppHandler)
+			r.Post("/embeddings", llamaCppHandler)
+			r.Post("/infill", llamaCppHandler)
+			r.Post("/metrics", llamaCppHandler)
+			r.Post("/props", llamaCppHandler)
+			r.Post("/reranking", llamaCppHandler)
+			r.Post("/tokenize", llamaCppHandler)
+
+			// OpenAI-compatible proxy endpoint
+			// Handles all POST requests to /v1/*, including:
+			//   - /v1/completions
+			//   - /v1/chat/completions
+			//   - /v1/embeddings
+			//   - /v1/rerank
+			//   - /v1/reranking
+			// llamaCppHandler is used here because some users of llama.cpp endpoints depend
+			// on "model" field being optional, and handler.OpenAIProxy requires it.
+			r.Post("/v1/*", llamaCppHandler)
+		})
+
+	})
+
 	// Serve WebUI files
 	if err := webui.SetupWebUI(r); err != nil {
 		fmt.Printf("Failed to set up WebUI: %v\n", err)