Support llama.cpp router mode for openai endpoints

pkg/backends/backend.go

@@ -14,6 +14,7 @@ const (
 	BackendTypeLlamaCpp BackendType = "llama_cpp"
 	BackendTypeMlxLm    BackendType = "mlx_lm"
 	BackendTypeVllm     BackendType = "vllm"
+	BackendTypeUnknown  BackendType = "unknown"
 )
 
 type backend interface {

pkg/instance/instance.go

@@ -7,6 +7,7 @@ import (
 	"net/http"
 	"time"
 
+	"llamactl/pkg/backends"
 	"llamactl/pkg/config"
 )
 
@@ -117,6 +118,14 @@ func (i *Instance) WaitForHealthy(timeout int) error {
 	return i.process.waitForHealthy(timeout)
 }
 
+func (i *Instance) GetBackendType() backends.BackendType {
+	opts := i.GetOptions()
+	if opts == nil {
+		return backends.BackendTypeUnknown
+	}
+	return opts.BackendOptions.BackendType
+}
+
 // GetOptions returns the current options
 func (i *Instance) GetOptions() *Options {
 	if i.options == nil {

pkg/instance/models.go

@@ -1,141 +0,0 @@
-package instance
-
-import (
-	"bytes"
-	"context"
-	"encoding/json"
-	"fmt"
-	"io"
-	"llamactl/pkg/backends"
-	"net/http"
-	"time"
-)
-
-// Model represents a model available in a llama.cpp instance
-type Model struct {
-	ID      string      `json:"id"`
-	Object  string      `json:"object"`
-	OwnedBy string      `json:"owned_by"`
-	Created int64       `json:"created"`
-	InCache bool        `json:"in_cache"`
-	Path    string      `json:"path"`
-	Status  ModelStatus `json:"status"`
-}
-
-// ModelStatus represents the status of a model in an instance
-type ModelStatus struct {
-	Value string   `json:"value"` // "loaded" | "loading" | "unloaded"
-	Args  []string `json:"args"`
-}
-
-// IsLlamaCpp checks if this instance is a llama.cpp instance
-func (i *Instance) IsLlamaCpp() bool {
-	opts := i.GetOptions()
-	if opts == nil {
-		return false
-	}
-	return opts.BackendOptions.BackendType == backends.BackendTypeLlamaCpp
-}
-
-// GetModels fetches the models available in this llama.cpp instance
-func (i *Instance) GetModels() ([]Model, error) {
-	if !i.IsLlamaCpp() {
-		return nil, fmt.Errorf("instance %s is not a llama.cpp instance", i.Name)
-	}
-
-	if !i.IsRunning() {
-		return nil, fmt.Errorf("instance %s is not running", i.Name)
-	}
-
-	var result struct {
-		Data []Model `json:"data"`
-	}
-	if err := i.doRequest("GET", "/models", nil, &result, 10*time.Second); err != nil {
-		return nil, fmt.Errorf("failed to fetch models: %w", err)
-	}
-
-	return result.Data, nil
-}
-
-// LoadModel loads a model in this llama.cpp instance
-func (i *Instance) LoadModel(modelName string) error {
-	if !i.IsLlamaCpp() {
-		return fmt.Errorf("instance %s is not a llama.cpp instance", i.Name)
-	}
-
-	if !i.IsRunning() {
-		return fmt.Errorf("instance %s is not running", i.Name)
-	}
-
-	// Make the load request
-	reqBody := map[string]string{"model": modelName}
-	if err := i.doRequest("POST", "/models/load", reqBody, nil, 30*time.Second); err != nil {
-		return fmt.Errorf("failed to load model: %w", err)
-	}
-
-	return nil
-}
-
-// UnloadModel unloads a model from this llama.cpp instance
-func (i *Instance) UnloadModel(modelName string) error {
-	if !i.IsLlamaCpp() {
-		return fmt.Errorf("instance %s is not a llama.cpp instance", i.Name)
-	}
-
-	if !i.IsRunning() {
-		return fmt.Errorf("instance %s is not running", i.Name)
-	}
-
-	// Make the unload request
-	reqBody := map[string]string{"model": modelName}
-	if err := i.doRequest("POST", "/models/unload", reqBody, nil, 30*time.Second); err != nil {
-		return fmt.Errorf("failed to unload model: %w", err)
-	}
-
-	return nil
-}
-
-// doRequest makes an HTTP request to this instance's backend
-func (i *Instance) doRequest(method, path string, reqBody, respBody any, timeout time.Duration) error {
-	url := fmt.Sprintf("http://%s:%d%s", i.GetHost(), i.GetPort(), path)
-
-	var bodyReader io.Reader
-	if reqBody != nil {
-		bodyBytes, err := json.Marshal(reqBody)
-		if err != nil {
-			return fmt.Errorf("failed to marshal request body: %w", err)
-		}
-		bodyReader = bytes.NewReader(bodyBytes)
-	}
-
-	ctx, cancel := context.WithTimeout(context.Background(), timeout)
-	defer cancel()
-
-	req, err := http.NewRequestWithContext(ctx, method, url, bodyReader)
-	if err != nil {
-		return fmt.Errorf("failed to create request: %w", err)
-	}
-
-	if reqBody != nil {
-		req.Header.Set("Content-Type", "application/json")
-	}
-
-	resp, err := http.DefaultClient.Do(req)
-	if err != nil {
-		return err
-	}
-	defer resp.Body.Close()
-
-	if resp.StatusCode != http.StatusOK {
-		bodyBytes, _ := io.ReadAll(resp.Body)
-		return fmt.Errorf("status %d: %s", resp.StatusCode, string(bodyBytes))
-	}
-
-	if respBody != nil {
-		if err := json.NewDecoder(resp.Body).Decode(respBody); err != nil {
-			return fmt.Errorf("failed to decode response: %w", err)
-		}
-	}
-
-	return nil
-}

pkg/server/handlers_backends.go

@@ -8,8 +8,6 @@ import (
 	"net/http"
 	"os/exec"
 	"strings"
-
-	"github.com/go-chi/chi/v5"
 )
 
 // ParseCommandRequest represents the request body for backend command parsing
@@ -322,24 +320,41 @@ func (h *Handler) LlamaServerListDevicesHandler() http.HandlerFunc {
 // @Router /api/v1/llama-cpp/{name}/models [get]
 func (h *Handler) LlamaCppListModels() http.HandlerFunc {
 	return func(w http.ResponseWriter, r *http.Request) {
-		inst, err := h.getInstance(r)
+		inst, err := h.validateLlamaCppInstance(r)
 		if err != nil {
-			writeError(w, http.StatusBadRequest, "invalid_instance", err.Error())
+			writeError(w, http.StatusBadRequest, "invalid instance", err.Error())
 			return
 		}
 
-		models, err := inst.GetModels()
-		if err != nil {
-			writeError(w, http.StatusBadRequest, "get_models_failed", err.Error())
+		// Check instance permissions
+		if err := h.authMiddleware.CheckInstancePermission(r.Context(), inst.ID); err != nil {
+			writeError(w, http.StatusForbidden, "permission_denied", err.Error())
 			return
 		}
 
-		response := map[string]any{
-			"object": "list",
-			"data":   models,
+		// Check if instance is shutting down before autostart logic
+		if inst.GetStatus() == instance.ShuttingDown {
+			writeError(w, http.StatusServiceUnavailable, "instance_shutting_down", "Instance is shutting down")
+			return
 		}
 
-		writeJSON(w, http.StatusOK, response)
+		if !inst.IsRemote() && !inst.IsRunning() {
+			err := h.ensureInstanceRunning(inst)
+			if err != nil {
+				writeError(w, http.StatusInternalServerError, "instance start failed", err.Error())
+				return
+			}
+		}
+
+		// Modify request path to /models for proxying
+		r.URL.Path = "/models"
+
+		// Use instance's ServeHTTP which tracks inflight requests and handles shutting down state
+		err = inst.ServeHTTP(w, r)
+		if err != nil {
+			// Error is already handled in ServeHTTP (response written)
+			return
+		}
 	}
 }
 
@@ -357,23 +372,41 @@ func (h *Handler) LlamaCppListModels() http.HandlerFunc {
 // @Router /api/v1/llama-cpp/{name}/models/{model}/load [post]
 func (h *Handler) LlamaCppLoadModel() http.HandlerFunc {
 	return func(w http.ResponseWriter, r *http.Request) {
-		inst, err := h.getInstance(r)
+		inst, err := h.validateLlamaCppInstance(r)
 		if err != nil {
-			writeError(w, http.StatusBadRequest, "invalid_instance", err.Error())
+			writeError(w, http.StatusBadRequest, "invalid instance", err.Error())
 			return
 		}
 
-		modelName := chi.URLParam(r, "model")
-
-		if err := inst.LoadModel(modelName); err != nil {
-			writeError(w, http.StatusBadRequest, "load_model_failed", err.Error())
+		// Check instance permissions
+		if err := h.authMiddleware.CheckInstancePermission(r.Context(), inst.ID); err != nil {
+			writeError(w, http.StatusForbidden, "permission_denied", err.Error())
 			return
 		}
 
-		writeJSON(w, http.StatusOK, map[string]string{
-			"status":  "success",
-			"message": fmt.Sprintf("Model %s loaded successfully", modelName),
-		})
+		// Check if instance is shutting down before autostart logic
+		if inst.GetStatus() == instance.ShuttingDown {
+			writeError(w, http.StatusServiceUnavailable, "instance_shutting_down", "Instance is shutting down")
+			return
+		}
+
+		if !inst.IsRemote() && !inst.IsRunning() {
+			err := h.ensureInstanceRunning(inst)
+			if err != nil {
+				writeError(w, http.StatusInternalServerError, "instance start failed", err.Error())
+				return
+			}
+		}
+
+		// Modify request path to /models/load for proxying
+		r.URL.Path = "/models/load"
+
+		// Use instance's ServeHTTP which tracks inflight requests and handles shutting down state
+		err = inst.ServeHTTP(w, r)
+		if err != nil {
+			// Error is already handled in ServeHTTP (response written)
+			return
+		}
 	}
 }
 
@@ -391,22 +424,40 @@ func (h *Handler) LlamaCppLoadModel() http.HandlerFunc {
 // @Router /api/v1/llama-cpp/{name}/models/{model}/unload [post]
 func (h *Handler) LlamaCppUnloadModel() http.HandlerFunc {
 	return func(w http.ResponseWriter, r *http.Request) {
-		inst, err := h.getInstance(r)
+		inst, err := h.validateLlamaCppInstance(r)
 		if err != nil {
-			writeError(w, http.StatusBadRequest, "invalid_instance", err.Error())
+			writeError(w, http.StatusBadRequest, "invalid instance", err.Error())
 			return
 		}
 
-		modelName := chi.URLParam(r, "model")
-
-		if err := inst.UnloadModel(modelName); err != nil {
-			writeError(w, http.StatusBadRequest, "unload_model_failed", err.Error())
+		// Check instance permissions
+		if err := h.authMiddleware.CheckInstancePermission(r.Context(), inst.ID); err != nil {
+			writeError(w, http.StatusForbidden, "permission_denied", err.Error())
 			return
 		}
 
-		writeJSON(w, http.StatusOK, map[string]string{
-			"status":  "success",
-			"message": fmt.Sprintf("Model %s unloaded successfully", modelName),
-		})
+		// Check if instance is shutting down before autostart logic
+		if inst.GetStatus() == instance.ShuttingDown {
+			writeError(w, http.StatusServiceUnavailable, "instance_shutting_down", "Instance is shutting down")
+			return
+		}
+
+		if !inst.IsRemote() && !inst.IsRunning() {
+			err := h.ensureInstanceRunning(inst)
+			if err != nil {
+				writeError(w, http.StatusInternalServerError, "instance start failed", err.Error())
+				return
+			}
+		}
+
+		// Modify request path to /models/unload for proxying
+		r.URL.Path = "/models/unload"
+
+		// Use instance's ServeHTTP which tracks inflight requests and handles shutting down state
+		err = inst.ServeHTTP(w, r)
+		if err != nil {
+			// Error is already handled in ServeHTTP (response written)
+			return
+		}
 	}
 }

pkg/server/handlers_openai.go

@@ -5,9 +5,11 @@ import (
 	"encoding/json"
 	"fmt"
 	"io"
+	"llamactl/pkg/backends"
 	"llamactl/pkg/instance"
 	"llamactl/pkg/validation"
 	"net/http"
+	"strings"
 )
 
 // OpenAIListInstancesResponse represents the response structure for listing instances (models) in OpenAI-compatible format
@@ -24,6 +26,53 @@ type OpenAIInstance struct {
 	OwnedBy string `json:"owned_by"`
 }
 
+// LlamaCppModel represents a model available in a llama.cpp instance
+type LlamaCppModel struct {
+	ID      string              `json:"id"`
+	Object  string              `json:"object"`
+	OwnedBy string              `json:"owned_by"`
+	Created int64               `json:"created"`
+	InCache bool                `json:"in_cache"`
+	Path    string              `json:"path"`
+	Status  LlamaCppModelStatus `json:"status"`
+}
+
+// LlamaCppModelStatus represents the status of a model in a llama.cpp instance
+type LlamaCppModelStatus struct {
+	Value string   `json:"value"` // "loaded" | "loading" | "unloaded"
+	Args  []string `json:"args"`
+}
+
+// fetchLlamaCppModels fetches models from a llama.cpp instance using the proxy
+func fetchLlamaCppModels(inst *instance.Instance) ([]LlamaCppModel, error) {
+	// Create a request to the instance's /models endpoint
+	req, err := http.NewRequest("GET", fmt.Sprintf("http://%s:%d/models", inst.GetHost(), inst.GetPort()), nil)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create request: %w", err)
+	}
+
+	// Use a custom response writer to capture the response
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		return nil, err
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode != http.StatusOK {
+		bodyBytes, _ := io.ReadAll(resp.Body)
+		return nil, fmt.Errorf("status %d: %s", resp.StatusCode, string(bodyBytes))
+	}
+
+	var result struct {
+		Data []LlamaCppModel `json:"data"`
+	}
+	if err := json.NewDecoder(resp.Body).Decode(&result); err != nil {
+		return nil, fmt.Errorf("failed to decode response: %w", err)
+	}
+
+	return result.Data, nil
+}
+
 // OpenAIListInstances godoc
 // @Summary List instances in OpenAI-compatible format
 // @Description Returns a list of instances in a format compatible with OpenAI API
@@ -46,9 +95,9 @@ func (h *Handler) OpenAIListInstances() http.HandlerFunc {
 		// For each llama.cpp instance, try to fetch models and add them as separate entries
 		for _, inst := range instances {
 
-			if inst.IsLlamaCpp() && inst.IsRunning() {
+			if inst.GetBackendType() == backends.BackendTypeLlamaCpp && inst.IsRunning() {
 				// Try to fetch models from the instance
-				models, err := inst.GetModels()
+				models, err := fetchLlamaCppModels(inst)
 				if err != nil {
 					fmt.Printf("Failed to fetch models from instance %s: %v", inst.Name, err)
 					continue
@@ -56,9 +105,9 @@ func (h *Handler) OpenAIListInstances() http.HandlerFunc {
 
 				for _, model := range models {
 					openaiInstances = append(openaiInstances, OpenAIInstance{
-						ID:      model.ID,
+						ID:      inst.Name + "/" + model.ID,
 						Object:  "model",
-						Created: model.Created,
+						Created: inst.Created,
 						OwnedBy: inst.Name,
 					})
 				}
@@ -115,17 +164,24 @@ func (h *Handler) OpenAIProxy() http.HandlerFunc {
 			return
 		}
 
-		modelName, ok := requestBody["model"].(string)
-		if !ok || modelName == "" {
+		reqModelName, ok := requestBody["model"].(string)
+		if !ok || reqModelName == "" {
 			writeError(w, http.StatusBadRequest, "invalid_request", "Model name is required")
 			return
 		}
 
-		// Resolve model name to instance name (checks instance names first, then model registry)
-		instanceName, err := h.InstanceManager.ResolveInstance(modelName)
-		if err != nil {
-			writeError(w, http.StatusBadRequest, "model_not_found", err.Error())
-			return
+		// Parse instance name and model name from <instance_name>/<model_name> format
+		var instanceName string
+		var modelName string
+
+		// Check if model name contains "/"
+		if idx := strings.Index(reqModelName, "/"); idx != -1 {
+			// Split into instance and model parts
+			instanceName = reqModelName[:idx]
+			modelName = reqModelName[idx+1:]
+		} else {
+			instanceName = reqModelName
+			modelName = reqModelName
 		}
 
 		// Validate instance name at the entry point
@@ -154,6 +210,11 @@ func (h *Handler) OpenAIProxy() http.HandlerFunc {
 			return
 		}
 
+		if inst.IsRemote() {
+			// Don't replace model name for remote instances
+			modelName = reqModelName
+		}
+
 		if !inst.IsRemote() && !inst.IsRunning() {
 			err := h.ensureInstanceRunning(inst)
 			if err != nil {
@@ -162,6 +223,16 @@ func (h *Handler) OpenAIProxy() http.HandlerFunc {
 			}
 		}
 
+		// Update the request body with just the model name
+		requestBody["model"] = modelName
+
+		// Re-marshal the updated body
+		bodyBytes, err = json.Marshal(requestBody)
+		if err != nil {
+			writeError(w, http.StatusInternalServerError, "marshal_error", "Failed to update request body")
+			return
+		}
+
 		// Recreate the request body from the bytes we read
 		r.Body = io.NopCloser(bytes.NewReader(bodyBytes))
 		r.ContentLength = int64(len(bodyBytes))