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/manager/manager.go

@@ -19,13 +19,11 @@ type InstanceManager interface {
 	UpdateInstance(name string, options *instance.Options) (*instance.Instance, error)
 	DeleteInstance(name string) error
 	StartInstance(name string) (*instance.Instance, error)
-	IsMaxRunningInstancesReached() bool
+	AtMaxRunning() bool
 	StopInstance(name string) (*instance.Instance, error)
 	EvictLRUInstance() error
 	RestartInstance(name string) (*instance.Instance, error)
 	GetInstanceLogs(name string, numLines int) (string, error)
-	ResolveInstance(modelName string) (string, error)
-	RefreshModelRegistry(inst *instance.Instance) error
 	Shutdown()
 }
 
@@ -36,7 +34,6 @@ type instanceManager struct {
 	db        database.InstanceStore
 	remote    *remoteManager
 	lifecycle *lifecycleManager
-	models    *modelRegistry
 
 	// Configuration
 	globalConfig *config.AppConfig
@@ -63,16 +60,12 @@ func New(globalConfig *config.AppConfig, db database.InstanceStore) InstanceMana
 	// Initialize remote manager
 	remote := newRemoteManager(globalConfig.Nodes, 30*time.Second)
 
-	// Initialize model registry
-	models := newModelRegistry()
-
 	// Create manager instance
 	im := &instanceManager{
 		registry:     registry,
 		ports:        ports,
 		db:           db,
 		remote:       remote,
-		models:       models,
 		globalConfig: globalConfig,
 	}
 
@@ -149,27 +142,9 @@ func (im *instanceManager) loadInstances() error {
 	// Auto-start instances that have auto-restart enabled
 	go im.autoStartInstances()
 
-	// Discover models from all running llama.cpp instances
-	go im.discoverAllModels()
-
 	return nil
 }
 
-// discoverAllModels discovers and registers models for all running llama.cpp instances
-func (im *instanceManager) discoverAllModels() {
-	instances := im.registry.listRunning()
-
-	for _, inst := range instances {
-		if !inst.IsLlamaCpp() {
-			continue
-		}
-
-		if err := im.RefreshModelRegistry(inst); err != nil {
-			log.Printf("Failed to discover models for instance %s: %v", inst.Name, err)
-		}
-	}
-}
-
 // loadInstance loads a single persisted instance and adds it to the registry
 func (im *instanceManager) loadInstance(persistedInst *instance.Instance) error {
 	name := persistedInst.Name

pkg/manager/model_registry.go

@@ -1,79 +0,0 @@
-package manager
-
-import (
-	"fmt"
-	"llamactl/pkg/instance"
-	"sync"
-)
-
-// modelRegistry maintains a global mapping of model names to instance names
-// for llama.cpp instances. Model names must be globally unique across all instances.
-type modelRegistry struct {
-	mu              sync.RWMutex
-	modelToInstance map[string]string   // model name → instance name
-	instanceModels  map[string][]string // instance name → model names
-}
-
-// newModelRegistry creates a new model registry
-func newModelRegistry() *modelRegistry {
-	return &modelRegistry{
-		modelToInstance: make(map[string]string),
-		instanceModels:  make(map[string][]string),
-	}
-}
-
-// registerModels registers models from an instance to the registry.
-// Skips models that conflict with other instances and returns a list of conflicts.
-func (mr *modelRegistry) registerModels(instanceName string, models []instance.Model) []string {
-	mr.mu.Lock()
-	defer mr.mu.Unlock()
-
-	// Unregister any existing models for this instance first
-	mr.removeModels(instanceName)
-
-	// Register models, skipping conflicts
-	var modelNames []string
-	var conflicts []string
-
-	for _, model := range models {
-		// Check if this model conflicts with another instance
-		if existingInstance, exists := mr.modelToInstance[model.ID]; exists && existingInstance != instanceName {
-			conflicts = append(conflicts, fmt.Sprintf("%s (already in %s)", model.ID, existingInstance))
-			continue // Skip this model
-		}
-
-		// Register the model
-		mr.modelToInstance[model.ID] = instanceName
-		modelNames = append(modelNames, model.ID)
-	}
-
-	mr.instanceModels[instanceName] = modelNames
-
-	return conflicts
-}
-
-// unregisterModels removes all models for an instance
-func (mr *modelRegistry) unregisterModels(instanceName string) {
-	mr.mu.Lock()
-	defer mr.mu.Unlock()
-	mr.removeModels(instanceName)
-}
-
-// removeModels removes all models for an instance (caller must hold lock)
-func (mr *modelRegistry) removeModels(instanceName string) {
-	if models, exists := mr.instanceModels[instanceName]; exists {
-		for _, modelName := range models {
-			delete(mr.modelToInstance, modelName)
-		}
-		delete(mr.instanceModels, instanceName)
-	}
-}
-
-// getModelInstance returns the instance name that hosts the given model
-func (mr *modelRegistry) getModelInstance(modelName string) (string, bool) {
-	mr.mu.RLock()
-	defer mr.mu.RUnlock()
-
-	instanceName, exists := mr.modelToInstance[modelName]
-	return instanceName, exists
-}

pkg/manager/operations.go

@@ -337,9 +337,6 @@ func (im *instanceManager) DeleteInstance(name string) error {
 	// Release port (use ReleaseByInstance for proper cleanup)
 	im.ports.releaseByInstance(name)
 
-	// Unregister models when instance is deleted
-	im.onInstanceStopped(name)
-
 	// Remove from registry
 	if err := im.registry.remove(name); err != nil {
 		return fmt.Errorf("failed to remove instance from registry: %w", err)
@@ -386,7 +383,7 @@ func (im *instanceManager) StartInstance(name string) (*instance.Instance, error
 	}
 
 	// Check max running instances limit for local instances only
-	if im.IsMaxRunningInstancesReached() {
+	if im.AtMaxRunning() {
 		return nil, MaxRunningInstancesError(fmt.Errorf("maximum number of running instances (%d) reached", im.globalConfig.Instances.MaxRunningInstances))
 	}
 
@@ -399,13 +396,10 @@ func (im *instanceManager) StartInstance(name string) (*instance.Instance, error
 		log.Printf("Warning: failed to persist instance %s: %v", name, err)
 	}
 
-	// Discover and register models for llama.cpp instances
-	go im.onInstanceStarted(name)
-
 	return inst, nil
 }
 
-func (im *instanceManager) IsMaxRunningInstancesReached() bool {
+func (im *instanceManager) AtMaxRunning() bool {
 	if im.globalConfig.Instances.MaxRunningInstances == -1 {
 		return false
 	}
@@ -461,9 +455,6 @@ func (im *instanceManager) StopInstance(name string) (*instance.Instance, error)
 		log.Printf("Warning: failed to persist instance %s: %v", name, err)
 	}
 
-	// Unregister models when instance stops
-	im.onInstanceStopped(name)
-
 	return inst, nil
 }
 
@@ -544,73 +535,3 @@ func (im *instanceManager) setPortInOptions(options *instance.Options, port int)
 func (im *instanceManager) EvictLRUInstance() error {
 	return im.lifecycle.evictLRU()
 }
-
-// ResolveInstance resolves a model name to an instance name.
-// Precedence: instance name > model registry
-func (im *instanceManager) ResolveInstance(modelName string) (string, error) {
-	// Check if it's an instance name first
-	if _, err := im.GetInstance(modelName); err == nil {
-		return modelName, nil
-	}
-
-	// Check if it's a model name in the registry
-	if instanceName, exists := im.models.getModelInstance(modelName); exists {
-		return instanceName, nil
-	}
-
-	return "", fmt.Errorf("model or instance '%s' not found", modelName)
-}
-
-// RefreshModelRegistry refreshes the model registry for the given instance
-func (im *instanceManager) RefreshModelRegistry(inst *instance.Instance) error {
-	if !inst.IsRunning() {
-		return fmt.Errorf("instance %s is not running", inst.Name)
-	}
-
-	// Fetch models from instance and register them
-	models, err := inst.GetModels()
-	if err != nil {
-		return fmt.Errorf("failed to fetch models: %w", err)
-	}
-
-	// Register models, skipping conflicts
-	conflicts := im.models.registerModels(inst.Name, models)
-	if len(conflicts) > 0 {
-		log.Printf("Warning: Model name conflicts for instance %s (skipped): %v", inst.Name, conflicts)
-	}
-
-	// Check if instance name shadows any model names
-	if otherInstance, exists := im.models.getModelInstance(inst.Name); exists && otherInstance != inst.Name {
-		log.Printf("Warning: Instance name '%s' shadows model name from instance '%s'", inst.Name, otherInstance)
-	}
-
-	return nil
-}
-
-// onInstanceStarted is called when an instance successfully starts and becomes healthy
-func (im *instanceManager) onInstanceStarted(name string) {
-	inst, err := im.GetInstance(name)
-	if err != nil {
-		log.Printf("Failed to get instance %s for model discovery: %v", name, err)
-		return
-	}
-
-	// Only discover models for llama.cpp instances
-	if !inst.IsLlamaCpp() {
-		return
-	}
-
-	if err := inst.WaitForHealthy(30); err != nil {
-		log.Printf("Instance %s not healthy, skipping model discovery: %v", name, err)
-		return
-	}
-
-	if err := im.RefreshModelRegistry(inst); err != nil {
-		log.Printf("Failed to discover models for instance %s: %v", name, err)
-	}
-}
-
-// onInstanceStopped is called when an instance stops or is deleted
-func (im *instanceManager) onInstanceStopped(name string) {
-	im.models.unregisterModels(name)
-}

pkg/server/handlers.go

@@ -96,7 +96,7 @@ func (h *Handler) ensureInstanceRunning(inst *instance.Instance) error {
 		return fmt.Errorf("instance is not running and on-demand start is not enabled")
 	}
 
-	if h.InstanceManager.IsMaxRunningInstancesReached() {
+	if h.InstanceManager.AtMaxRunning() {
 		if h.cfg.Instances.EnableLRUEviction {
 			err := h.InstanceManager.EvictLRUInstance()
 			if err != nil {

pkg/server/handlers_backends.go

@@ -5,12 +5,9 @@ import (
 	"fmt"
 	"llamactl/pkg/backends"
 	"llamactl/pkg/instance"
-	"log"
 	"net/http"
 	"os/exec"
 	"strings"
-
-	"github.com/go-chi/chi/v5"
 )
 
 // ParseCommandRequest represents the request body for backend command parsing
@@ -323,24 +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
+		}
 	}
 }
 
@@ -358,28 +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
 		}
 
-		// Refresh the model registry
-		if err := h.InstanceManager.RefreshModelRegistry(inst); err != nil {
-			log.Printf("Warning: failed to refresh model registry after load: %v", err)
+		// 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, map[string]string{
-			"status":  "success",
-			"message": fmt.Sprintf("Model %s loaded successfully", modelName),
-		})
+		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
+		}
 	}
 }
 
@@ -397,27 +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
 		}
 
-		// Refresh the model registry
-		if err := h.InstanceManager.RefreshModelRegistry(inst); err != nil {
-			log.Printf("Warning: failed to refresh model registry after unload: %v", err)
+		// 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, map[string]string{
-			"status":  "success",
-			"message": fmt.Sprintf("Model %s unloaded successfully", modelName),
-		})
+		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))

test_llm.py

@@ -0,0 +1,74 @@
+#!/usr/bin/env python3
+"""
+Simple Python script to interact with local LLM server's OpenAI-compatible API
+"""
+
+import requests
+
+# Local LLM server configuration
+LLM_SERVER_URL = "http://localhost:8080/v1/chat/completions"
+MODEL_NAME = "proxy-test"  # Default model name, can be changed based on your setup
+
+def send_message(message, model=MODEL_NAME, temperature=0.7, max_tokens=1000):
+    """
+    Send a message to local LLM server API
+    
+    Args:
+        message (str): The message to send
+        model (str): Model name (depends on your LLM server setup)
+        temperature (float): Controls randomness (0.0 to 1.0)
+        max_tokens (int): Maximum tokens in response
+    
+    Returns:
+        str: The AI response or error message
+    """
+    
+    headers = {
+        "Content-Type": "application/json",
+        "Authorization": "Bearer test-inf"
+    }
+    
+    data = {
+        "model": model,
+        "messages": [
+            {
+                "role": "user",
+                "content": message
+            }
+        ],
+        "temperature": temperature,
+        "max_tokens": max_tokens,
+        "stream": False
+    }
+    
+    response = requests.post(LLM_SERVER_URL, headers=headers, json=data, timeout=60)
+    response.raise_for_status()
+    
+    result = response.json()
+    return result["choices"][0]["message"]["content"]
+
+def main():
+    """Run in interactive mode for continuous conversation"""
+    print("Local LLM Chat Client")
+    print("-" * 40)
+    
+    while True:
+        try:
+            user_input = input("\nYou: ").strip()
+                
+            if not user_input:
+                continue
+                
+            print("AI: ", end="", flush=True)
+            response = send_message(user_input)
+            print(response)
+            
+        except KeyboardInterrupt:
+            print("\nGoodbye!")
+            break
+        except EOFError:
+            print("\nGoodbye!")
+            break
+
+if __name__ == "__main__":
+    main()