server: usage api

Add a new /api/usage endpoint that shows aggregate usage statistics per model since the server started.

api/types.go

@@ -912,6 +912,19 @@ type UserResponse struct {
 	Plan      string    `json:"plan,omitempty"`
 }
 
+type UsageResponse struct {
+	// Start is the time the server started tracking usage (UTC, RFC 3339).
+	Start time.Time        `json:"start"`
+	Usage []ModelUsageData `json:"usage"`
+}
+
+type ModelUsageData struct {
+	Model            string `json:"model"`
+	Requests         int64  `json:"requests"`
+	PromptTokens     int64  `json:"prompt_tokens"`
+	CompletionTokens int64  `json:"completion_tokens"`
+}
+
 // Tensor describes the metadata for a given tensor.
 type Tensor struct {
 	Name  string   `json:"name"`

docs/api.md

@@ -15,6 +15,7 @@
 - [Push a Model](#push-a-model)
 - [Generate Embeddings](#generate-embeddings)
 - [List Running Models](#list-running-models)
+- [Usage](#usage)
 - [Version](#version)
 - [Experimental: Image Generation](#image-generation-experimental)
 
@@ -1854,6 +1855,53 @@ curl http://localhost:11434/api/embeddings -d '{
 }
 ```
 
+## Usage
+
+```
+GET /api/usage
+```
+
+Show aggregate usage statistics per model since the server started. All timestamps are UTC in RFC 3339 format.
+
+### Examples
+
+#### Request
+
+```shell
+curl http://localhost:11434/api/usage
+```
+
+#### Response
+
+```json
+{
+  "start": "2025-01-27T20:00:00Z",
+  "usage": [
+    {
+      "model": "llama3.2",
+      "requests": 5,
+      "prompt_tokens": 130,
+      "completion_tokens": 890
+    },
+    {
+      "model": "deepseek-r1",
+      "requests": 2,
+      "prompt_tokens": 48,
+      "completion_tokens": 312
+    }
+  ]
+}
+```
+
+#### Response fields
+
+- `start`: when the server started tracking usage (UTC, RFC 3339)
+- `usage`: list of per-model usage statistics
+  - `model`: model name
+  - `requests`: total number of completed requests
+  - `prompt_tokens`: total prompt tokens evaluated
+  - `completion_tokens`: total completion tokens generated
+
 ## Version
 
 ```

envconfig/config.go

@@ -201,7 +201,7 @@ var (
 	// Enable the new Ollama engine
 	NewEngine = Bool("OLLAMA_NEW_ENGINE")
 	// ContextLength sets the default context length
-	ContextLength = Uint("OLLAMA_CONTEXT_LENGTH", 0)
+	ContextLength = Uint("OLLAMA_CONTEXT_LENGTH", 4096)
 	// Auth enables authentication between the Ollama client and server
 	UseAuth = Bool("OLLAMA_AUTH")
 	// Enable Vulkan backend

envconfig/config_test.go

@@ -282,7 +282,7 @@ func TestVar(t *testing.T) {
 
 func TestContextLength(t *testing.T) {
 	cases := map[string]uint{
-		"":     0,
+		"":     4096,
 		"2048": 2048,
 	}
 

llm/server.go

@@ -80,7 +80,6 @@ type LlamaServer interface {
 	GetPort() int
 	GetDeviceInfos(ctx context.Context) []ml.DeviceInfo
 	HasExited() bool
-	ContextLength() int
 }
 
 // llmServer is an instance of a runner hosting a single model
@@ -1902,10 +1901,6 @@ func (s *llmServer) VRAMByGPU(id ml.DeviceID) uint64 {
 	return 0
 }
 
-func (s *llmServer) ContextLength() int {
-	return s.options.NumCtx
-}
-
 func (s *ollamaServer) GetDeviceInfos(ctx context.Context) []ml.DeviceInfo {
 	devices, err := ml.GetDevicesFromRunner(ctx, s)
 	if err != nil {

server/routes.go

@@ -75,12 +75,17 @@ func experimentEnabled(name string) bool {
 
 var useClient2 = experimentEnabled("client2")
 
+// Low VRAM mode is based on the sum of total VRAM (not free) and triggers
+// reduced context length on some models
+var lowVRAMThreshold uint64 = 20 * format.GibiByte
+
 var mode string = gin.DebugMode
 
 type Server struct {
-	addr          net.Addr
-	sched         *Scheduler
-	defaultNumCtx int
+	addr    net.Addr
+	sched   *Scheduler
+	lowVRAM bool
+	usage   *UsageTracker
 }
 
 func init() {
@@ -103,12 +108,8 @@ var (
 	errBadTemplate = errors.New("template error")
 )
 
-func (s *Server) modelOptions(model *Model, requestOpts map[string]any) (api.Options, error) {
+func modelOptions(model *Model, requestOpts map[string]any) (api.Options, error) {
 	opts := api.DefaultOptions()
-	if opts.NumCtx == 0 {
-		opts.NumCtx = s.defaultNumCtx
-	}
-
 	if err := opts.FromMap(model.Options); err != nil {
 		return api.Options{}, err
 	}
@@ -140,11 +141,20 @@ func (s *Server) scheduleRunner(ctx context.Context, name string, caps []model.C
 		return nil, nil, nil, fmt.Errorf("%s %w", name, err)
 	}
 
-	opts, err := s.modelOptions(model, requestOpts)
+	opts, err := modelOptions(model, requestOpts)
 	if err != nil {
 		return nil, nil, nil, err
 	}
 
+	// This model is much more capable with a larger context, so set that
+	// unless it would penalize performance too much
+	if !s.lowVRAM && slices.Contains([]string{
+		"gptoss", "gpt-oss",
+		"qwen3vl", "qwen3vlmoe",
+	}, model.Config.ModelFamily) {
+		opts.NumCtx = max(opts.NumCtx, 8192)
+	}
+
 	runnerCh, errCh := s.sched.GetRunner(ctx, model, opts, keepAlive)
 	var runner *runnerRef
 	select {
@@ -264,6 +274,10 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		c.Header("Content-Type", contentType)
 
 		fn := func(resp api.GenerateResponse) error {
+			if resp.Done {
+				s.usage.Record(origModel, resp.PromptEvalCount, resp.EvalCount)
+			}
+
 			resp.Model = origModel
 			resp.RemoteModel = m.Config.RemoteModel
 			resp.RemoteHost = m.Config.RemoteHost
@@ -570,6 +584,8 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 					}
 					res.Context = tokens
 				}
+
+				s.usage.Record(req.Model, cr.PromptEvalCount, cr.EvalCount)
 			}
 
 			if builtinParser != nil {
@@ -1581,6 +1597,8 @@ func (s *Server) GenerateRoutes(rc *ollama.Registry) (http.Handler, error) {
 	r.HEAD("/api/blobs/:digest", s.HeadBlobHandler)
 	r.POST("/api/copy", s.CopyHandler)
 
+	r.GET("/api/usage", s.UsageHandler)
+
 	// Inference
 	r.GET("/api/ps", s.PsHandler)
 	r.POST("/api/generate", s.GenerateHandler)
@@ -1649,7 +1667,7 @@ func Serve(ln net.Listener) error {
 		}
 	}
 
-	s := &Server{addr: ln.Addr()}
+	s := &Server{addr: ln.Addr(), usage: NewUsageTracker()}
 
 	var rc *ollama.Registry
 	if useClient2 {
@@ -1711,18 +1729,10 @@ func Serve(ln net.Listener) error {
 	for _, gpu := range gpus {
 		totalVRAM += gpu.TotalMemory - envconfig.GpuOverhead()
 	}
-
-	// Set default context based on VRAM tier
-	// Use slightly lower thresholds (47/23 GiB vs. 48/24 GiB) to account for small differences in the exact value
-	switch {
-	case totalVRAM >= 47*format.GibiByte:
-		s.defaultNumCtx = 262144
-	case totalVRAM >= 23*format.GibiByte:
-		s.defaultNumCtx = 32768
-	default:
-		s.defaultNumCtx = 4096
+	if totalVRAM < lowVRAMThreshold {
+		s.lowVRAM = true
+		slog.Info("entering low vram mode", "total vram", format.HumanBytes2(totalVRAM), "threshold", format.HumanBytes2(lowVRAMThreshold))
 	}
-	slog.Info("vram-based default context", "total_vram", format.HumanBytes2(totalVRAM), "default_num_ctx", s.defaultNumCtx)
 
 	err = srvr.Serve(ln)
 	// If server is closed from the signal handler, wait for the ctx to be done
@@ -1874,6 +1884,10 @@ func (s *Server) SignoutHandler(c *gin.Context) {
 	c.JSON(http.StatusOK, nil)
 }
 
+func (s *Server) UsageHandler(c *gin.Context) {
+	c.JSON(http.StatusOK, s.usage.Stats())
+}
+
 func (s *Server) PsHandler(c *gin.Context) {
 	models := []api.ProcessModelResponse{}
 
@@ -1896,8 +1910,8 @@ func (s *Server) PsHandler(c *gin.Context) {
 			Details:   modelDetails,
 			ExpiresAt: v.expiresAt,
 		}
-		if v.llama != nil {
-			mr.ContextLength = v.llama.ContextLength()
+		if v.Options != nil {
+			mr.ContextLength = v.Options.NumCtx
 		}
 		// The scheduler waits to set expiresAt, so if a model is loading it's
 		// possible that it will be set to the unix epoch. For those cases, just
@@ -2032,6 +2046,10 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		c.Header("Content-Type", contentType)
 
 		fn := func(resp api.ChatResponse) error {
+			if resp.Done {
+				s.usage.Record(origModel, resp.PromptEvalCount, resp.EvalCount)
+			}
+
 			resp.Model = origModel
 			resp.RemoteModel = m.Config.RemoteModel
 			resp.RemoteHost = m.Config.RemoteHost
@@ -2252,6 +2270,8 @@ func (s *Server) ChatHandler(c *gin.Context) {
 					res.DoneReason = r.DoneReason.String()
 					res.TotalDuration = time.Since(checkpointStart)
 					res.LoadDuration = checkpointLoaded.Sub(checkpointStart)
+
+					s.usage.Record(req.Model, r.PromptEvalCount, r.EvalCount)
 				}
 
 				if builtinParser != nil {

server/routes_debug_test.go

@@ -15,7 +15,6 @@ import (
 )
 
 func TestGenerateDebugRenderOnly(t *testing.T) {
-	t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
 	gin.SetMode(gin.TestMode)
 
 	mock := mockRunner{
@@ -209,7 +208,6 @@ func TestGenerateDebugRenderOnly(t *testing.T) {
 }
 
 func TestChatDebugRenderOnly(t *testing.T) {
-	t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
 	gin.SetMode(gin.TestMode)
 
 	mock := mockRunner{

server/routes_generate_renderer_test.go

@@ -20,7 +20,6 @@ import (
 // TestGenerateWithBuiltinRenderer tests that api/generate uses built-in renderers
 // when in chat-like flow (messages present, no suffix, no template)
 func TestGenerateWithBuiltinRenderer(t *testing.T) {
-	t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
 	gin.SetMode(gin.TestMode)
 
 	mock := mockRunner{
@@ -205,7 +204,6 @@ func TestGenerateWithBuiltinRenderer(t *testing.T) {
 
 // TestGenerateWithDebugRenderOnly tests that debug_render_only works with built-in renderers
 func TestGenerateWithDebugRenderOnly(t *testing.T) {
-	t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
 	gin.SetMode(gin.TestMode)
 
 	mock := mockRunner{

server/routes_generate_test.go

@@ -88,19 +88,39 @@ func TestGenerateChatRemote(t *testing.T) {
 		if r.Method != http.MethodPost {
 			t.Errorf("Expected POST request, got %s", r.Method)
 		}
-		if r.URL.Path != "/api/chat" {
-			t.Errorf("Expected path '/api/chat', got %s", r.URL.Path)
-		}
 
 		w.WriteHeader(http.StatusOK)
 		w.Header().Set("Content-Type", "application/json")
-		resp := api.ChatResponse{
-			Model:      "test",
-			Done:       true,
-			DoneReason: "load",
-		}
-		if err := json.NewEncoder(w).Encode(&resp); err != nil {
-			t.Fatal(err)
+
+		switch r.URL.Path {
+		case "/api/chat":
+			resp := api.ChatResponse{
+				Model:      "test",
+				Done:       true,
+				DoneReason: "load",
+				Metrics: api.Metrics{
+					PromptEvalCount: 10,
+					EvalCount:       20,
+				},
+			}
+			if err := json.NewEncoder(w).Encode(&resp); err != nil {
+				t.Fatal(err)
+			}
+		case "/api/generate":
+			resp := api.GenerateResponse{
+				Model:      "test",
+				Done:       true,
+				DoneReason: "stop",
+				Metrics: api.Metrics{
+					PromptEvalCount: 5,
+					EvalCount:       15,
+				},
+			}
+			if err := json.NewEncoder(w).Encode(&resp); err != nil {
+				t.Fatal(err)
+			}
+		default:
+			t.Errorf("unexpected path %s", r.URL.Path)
 		}
 	}))
 	defer rs.Close()
@@ -111,7 +131,7 @@ func TestGenerateChatRemote(t *testing.T) {
 	}
 
 	t.Setenv("OLLAMA_REMOTES", p.Hostname())
-	s := Server{}
+	s := Server{usage: NewUsageTracker()}
 	w := createRequest(t, s.CreateHandler, api.CreateRequest{
 		Model:      "test-cloud",
 		RemoteHost: rs.URL,
@@ -159,10 +179,64 @@ func TestGenerateChatRemote(t *testing.T) {
 			t.Errorf("expected done reason load, got %s", actual.DoneReason)
 		}
 	})
+
+	t.Run("remote chat usage tracking", func(t *testing.T) {
+		stats := s.usage.Stats()
+		found := false
+		for _, m := range stats.Usage {
+			if m.Model == "test-cloud" {
+				found = true
+				if m.Requests != 1 {
+					t.Errorf("expected 1 request, got %d", m.Requests)
+				}
+				if m.PromptTokens != 10 {
+					t.Errorf("expected 10 prompt tokens, got %d", m.PromptTokens)
+				}
+				if m.CompletionTokens != 20 {
+					t.Errorf("expected 20 completion tokens, got %d", m.CompletionTokens)
+				}
+			}
+		}
+		if !found {
+			t.Error("expected usage entry for test-cloud")
+		}
+	})
+
+	t.Run("remote generate usage tracking", func(t *testing.T) {
+		// Reset the tracker for a clean test
+		s.usage = NewUsageTracker()
+
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test-cloud",
+			Prompt: "hello",
+		})
+		if w.Code != http.StatusOK {
+			t.Fatalf("expected status 200, got %d", w.Code)
+		}
+
+		stats := s.usage.Stats()
+		found := false
+		for _, m := range stats.Usage {
+			if m.Model == "test-cloud" {
+				found = true
+				if m.Requests != 1 {
+					t.Errorf("expected 1 request, got %d", m.Requests)
+				}
+				if m.PromptTokens != 5 {
+					t.Errorf("expected 5 prompt tokens, got %d", m.PromptTokens)
+				}
+				if m.CompletionTokens != 15 {
+					t.Errorf("expected 15 completion tokens, got %d", m.CompletionTokens)
+				}
+			}
+		}
+		if !found {
+			t.Error("expected usage entry for test-cloud")
+		}
+	})
 }
 
 func TestGenerateChat(t *testing.T) {
-	t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
 	gin.SetMode(gin.TestMode)
 
 	mock := mockRunner{
@@ -879,7 +953,6 @@ func TestGenerateChat(t *testing.T) {
 }
 
 func TestGenerate(t *testing.T) {
-	t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
 	gin.SetMode(gin.TestMode)
 
 	mock := mockRunner{
@@ -2357,7 +2430,6 @@ func TestGenerateWithImages(t *testing.T) {
 // TestImageGenerateStreamFalse tests that image generation respects stream=false
 // and returns a single JSON response instead of streaming ndjson.
 func TestImageGenerateStreamFalse(t *testing.T) {
-	t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
 	gin.SetMode(gin.TestMode)
 
 	p := t.TempDir()

server/routes_options_test.go

@@ -1,127 +0,0 @@
-package server
-
-import (
-	"testing"
-)
-
-func TestModelOptionsNumCtxPriority(t *testing.T) {
-	tests := []struct {
-		name           string
-		envContextLen  string // empty means not set (uses 0 sentinel)
-		defaultNumCtx  int    // VRAM-based default
-		modelNumCtx    int    // 0 means not set in model
-		requestNumCtx  int    // 0 means not set in request
-		expectedNumCtx int
-	}{
-		{
-			name:           "vram default when nothing else set",
-			envContextLen:  "",
-			defaultNumCtx:  32768,
-			modelNumCtx:    0,
-			requestNumCtx:  0,
-			expectedNumCtx: 32768,
-		},
-		{
-			name:           "env var overrides vram default",
-			envContextLen:  "8192",
-			defaultNumCtx:  32768,
-			modelNumCtx:    0,
-			requestNumCtx:  0,
-			expectedNumCtx: 8192,
-		},
-		{
-			name:           "model overrides vram default",
-			envContextLen:  "",
-			defaultNumCtx:  32768,
-			modelNumCtx:    16384,
-			requestNumCtx:  0,
-			expectedNumCtx: 16384,
-		},
-		{
-			name:           "model overrides env var",
-			envContextLen:  "8192",
-			defaultNumCtx:  32768,
-			modelNumCtx:    16384,
-			requestNumCtx:  0,
-			expectedNumCtx: 16384,
-		},
-		{
-			name:           "request overrides everything",
-			envContextLen:  "8192",
-			defaultNumCtx:  32768,
-			modelNumCtx:    16384,
-			requestNumCtx:  4096,
-			expectedNumCtx: 4096,
-		},
-		{
-			name:           "request overrides vram default",
-			envContextLen:  "",
-			defaultNumCtx:  32768,
-			modelNumCtx:    0,
-			requestNumCtx:  4096,
-			expectedNumCtx: 4096,
-		},
-		{
-			name:           "request overrides model",
-			envContextLen:  "",
-			defaultNumCtx:  32768,
-			modelNumCtx:    16384,
-			requestNumCtx:  4096,
-			expectedNumCtx: 4096,
-		},
-		{
-			name:           "low vram tier default",
-			envContextLen:  "",
-			defaultNumCtx:  4096,
-			modelNumCtx:    0,
-			requestNumCtx:  0,
-			expectedNumCtx: 4096,
-		},
-		{
-			name:           "high vram tier default",
-			envContextLen:  "",
-			defaultNumCtx:  262144,
-			modelNumCtx:    0,
-			requestNumCtx:  0,
-			expectedNumCtx: 262144,
-		},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			// Set or clear environment variable
-			if tt.envContextLen != "" {
-				t.Setenv("OLLAMA_CONTEXT_LENGTH", tt.envContextLen)
-			}
-
-			// Create server with VRAM-based default
-			s := &Server{
-				defaultNumCtx: tt.defaultNumCtx,
-			}
-
-			// Create model options (use float64 as FromMap expects JSON-style numbers)
-			var modelOpts map[string]any
-			if tt.modelNumCtx != 0 {
-				modelOpts = map[string]any{"num_ctx": float64(tt.modelNumCtx)}
-			}
-			model := &Model{
-				Options: modelOpts,
-			}
-
-			// Create request options (use float64 as FromMap expects JSON-style numbers)
-			var requestOpts map[string]any
-			if tt.requestNumCtx != 0 {
-				requestOpts = map[string]any{"num_ctx": float64(tt.requestNumCtx)}
-			}
-
-			opts, err := s.modelOptions(model, requestOpts)
-			if err != nil {
-				t.Fatalf("modelOptions failed: %v", err)
-			}
-
-			if opts.NumCtx != tt.expectedNumCtx {
-				t.Errorf("NumCtx = %d, want %d", opts.NumCtx, tt.expectedNumCtx)
-			}
-		})
-	}
-}

server/sched_test.go

@@ -804,7 +804,6 @@ func (s *mockLlm) GetPort() int                                       { return -
 func (s *mockLlm) GetDeviceInfos(ctx context.Context) []ml.DeviceInfo { return nil }
 func (s *mockLlm) HasExited() bool                                    { return false }
 func (s *mockLlm) GetActiveDeviceIDs() []ml.DeviceID                  { return nil }
-func (s *mockLlm) ContextLength() int                                 { return 0 }
 
 // TestImageGenRunnerCanBeEvicted verifies that an image generation model
 // loaded in the scheduler can be evicted when idle.

server/usage.go

@@ -0,0 +1,62 @@
+package server
+
+import (
+	"sync"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+type ModelUsage struct {
+	Requests         int64
+	PromptTokens     int64
+	CompletionTokens int64
+}
+
+type UsageTracker struct {
+	mu     sync.Mutex
+	start  time.Time
+	models map[string]*ModelUsage
+}
+
+func NewUsageTracker() *UsageTracker {
+	return &UsageTracker{
+		start:  time.Now().UTC(),
+		models: make(map[string]*ModelUsage),
+	}
+}
+
+func (u *UsageTracker) Record(model string, promptTokens, completionTokens int) {
+	u.mu.Lock()
+	defer u.mu.Unlock()
+
+	m, ok := u.models[model]
+	if !ok {
+		m = &ModelUsage{}
+		u.models[model] = m
+	}
+
+	m.Requests++
+	m.PromptTokens += int64(promptTokens)
+	m.CompletionTokens += int64(completionTokens)
+}
+
+func (u *UsageTracker) Stats() api.UsageResponse {
+	u.mu.Lock()
+	defer u.mu.Unlock()
+
+	byModel := make([]api.ModelUsageData, 0, len(u.models))
+	for model, usage := range u.models {
+		byModel = append(byModel, api.ModelUsageData{
+			Model:            model,
+			Requests:         usage.Requests,
+			PromptTokens:     usage.PromptTokens,
+			CompletionTokens: usage.CompletionTokens,
+		})
+	}
+
+	return api.UsageResponse{
+		Start: u.start,
+		Usage: byModel,
+	}
+}

server/usage_test.go

@@ -0,0 +1,136 @@
+package server
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"sync"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/ollama/ollama/api"
+)
+
+func TestUsageTrackerRecord(t *testing.T) {
+	tracker := NewUsageTracker()
+
+	tracker.Record("model-a", 10, 20)
+	tracker.Record("model-a", 5, 15)
+	tracker.Record("model-b", 100, 200)
+
+	stats := tracker.Stats()
+
+	if len(stats.Usage) != 2 {
+		t.Fatalf("expected 2 models, got %d", len(stats.Usage))
+	}
+
+	lookup := make(map[string]api.ModelUsageData)
+	for _, m := range stats.Usage {
+		lookup[m.Model] = m
+	}
+
+	a := lookup["model-a"]
+	if a.Requests != 2 {
+		t.Errorf("model-a requests: expected 2, got %d", a.Requests)
+	}
+	if a.PromptTokens != 15 {
+		t.Errorf("model-a prompt tokens: expected 15, got %d", a.PromptTokens)
+	}
+	if a.CompletionTokens != 35 {
+		t.Errorf("model-a completion tokens: expected 35, got %d", a.CompletionTokens)
+	}
+
+	b := lookup["model-b"]
+	if b.Requests != 1 {
+		t.Errorf("model-b requests: expected 1, got %d", b.Requests)
+	}
+	if b.PromptTokens != 100 {
+		t.Errorf("model-b prompt tokens: expected 100, got %d", b.PromptTokens)
+	}
+	if b.CompletionTokens != 200 {
+		t.Errorf("model-b completion tokens: expected 200, got %d", b.CompletionTokens)
+	}
+}
+
+func TestUsageTrackerConcurrent(t *testing.T) {
+	tracker := NewUsageTracker()
+
+	var wg sync.WaitGroup
+	for i := 0; i < 100; i++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			tracker.Record("model-a", 1, 2)
+		}()
+	}
+	wg.Wait()
+
+	stats := tracker.Stats()
+	if len(stats.Usage) != 1 {
+		t.Fatalf("expected 1 model, got %d", len(stats.Usage))
+	}
+
+	m := stats.Usage[0]
+	if m.Requests != 100 {
+		t.Errorf("requests: expected 100, got %d", m.Requests)
+	}
+	if m.PromptTokens != 100 {
+		t.Errorf("prompt tokens: expected 100, got %d", m.PromptTokens)
+	}
+	if m.CompletionTokens != 200 {
+		t.Errorf("completion tokens: expected 200, got %d", m.CompletionTokens)
+	}
+}
+
+func TestUsageTrackerStart(t *testing.T) {
+	tracker := NewUsageTracker()
+
+	stats := tracker.Stats()
+	if stats.Start.IsZero() {
+		t.Error("expected non-zero start time")
+	}
+}
+
+func TestUsageHandler(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	s := &Server{
+		usage: NewUsageTracker(),
+	}
+
+	s.usage.Record("llama3", 50, 100)
+	s.usage.Record("llama3", 25, 50)
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest(http.MethodGet, "/api/usage", nil)
+
+	s.UsageHandler(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	var resp api.UsageResponse
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("failed to unmarshal response: %v", err)
+	}
+
+	if len(resp.Usage) != 1 {
+		t.Fatalf("expected 1 model, got %d", len(resp.Usage))
+	}
+
+	m := resp.Usage[0]
+	if m.Model != "llama3" {
+		t.Errorf("expected model llama3, got %s", m.Model)
+	}
+	if m.Requests != 2 {
+		t.Errorf("expected 2 requests, got %d", m.Requests)
+	}
+	if m.PromptTokens != 75 {
+		t.Errorf("expected 75 prompt tokens, got %d", m.PromptTokens)
+	}
+	if m.CompletionTokens != 150 {
+		t.Errorf("expected 150 completion tokens, got %d", m.CompletionTokens)
+	}
+}

x/imagegen/server.go

@@ -347,11 +347,6 @@ func (s *Server) VRAMByGPU(id ml.DeviceID) uint64 {
 	return s.vramSize
 }
 
-// Context length is not applicable for image generation.
-func (s *Server) ContextLength() int {
-	return 0
-}
-
 func (s *Server) Embedding(ctx context.Context, input string) ([]float32, int, error) {
 	return nil, 0, errors.New("not supported")
 }