x/imagegen: clean up image generation code (#13725)

server/routes.go

@@ -51,7 +51,6 @@ import (
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
 	"github.com/ollama/ollama/x/imagegen"
-	imagegenapi "github.com/ollama/ollama/x/imagegen/api"
 )
 
 const signinURLStr = "https://ollama.com/connect?name=%s&key=%s"
@@ -164,29 +163,6 @@ func (s *Server) scheduleRunner(ctx context.Context, name string, caps []model.C
 	return runner.llama, model, &opts, nil
 }
 
-// ScheduleImageGenRunner schedules an image generation model runner.
-// This implements the imagegenapi.RunnerScheduler interface.
-func (s *Server) ScheduleImageGenRunner(c *gin.Context, modelName string, opts api.Options, keepAlive *api.Duration) (llm.LlamaServer, error) {
-	m := &Model{
-		Name:      modelName,
-		ShortName: modelName,
-		ModelPath: modelName, // For image gen, ModelPath is just the model name
-		Config: model.ConfigV2{
-			Capabilities: []string{"image"},
-		},
-	}
-
-	runnerCh, errCh := s.sched.GetRunner(c.Request.Context(), m, opts, keepAlive)
-	var runner *runnerRef
-	select {
-	case runner = <-runnerCh:
-	case err := <-errCh:
-		return nil, err
-	}
-
-	return runner.llama, nil
-}
-
 func signinURL() (string, error) {
 	pubKey, err := auth.GetPublicKey()
 	if err != nil {
@@ -214,12 +190,6 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}
 
-	// Check if this is a known image generation model
-	if imagegen.ResolveModelName(req.Model) != "" {
-		imagegenapi.HandleGenerateRequest(c, s, req.Model, req.Prompt, req.KeepAlive, streamResponse)
-		return
-	}
-
 	name := model.ParseName(req.Model)
 	if !name.IsValid() {
 		// Ideally this is "invalid model name" but we're keeping with
@@ -1587,13 +1557,12 @@ func (s *Server) GenerateRoutes(rc *ollama.Registry) (http.Handler, error) {
 	r.GET("/v1/models", middleware.ListMiddleware(), s.ListHandler)
 	r.GET("/v1/models/:model", middleware.RetrieveMiddleware(), s.ShowHandler)
 	r.POST("/v1/responses", middleware.ResponsesMiddleware(), s.ChatHandler)
+	// Experimental OpenAI-compatible image generation endpoint
+	r.POST("/v1/images/generations", s.handleImageGeneration)
 
 	// Inference (Anthropic compatibility)
 	r.POST("/v1/messages", middleware.AnthropicMessagesMiddleware(), s.ChatHandler)
 
-	// Experimental image generation support
-	imagegenapi.RegisterRoutes(r, s)
-
 	if rc != nil {
 		// wrap old with new
 		rs := &registry.Local{
@@ -1911,6 +1880,62 @@ func toolCallId() string {
 	return "call_" + strings.ToLower(string(b))
 }
 
+func (s *Server) handleImageGeneration(c *gin.Context) {
+	var req struct {
+		Model  string `json:"model"`
+		Prompt string `json:"prompt"`
+		Size   string `json:"size"`
+	}
+	if err := c.ShouldBindJSON(&req); err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+
+	m, err := GetModel(req.Model)
+	if err != nil {
+		c.JSON(http.StatusNotFound, gin.H{"error": err.Error()})
+		return
+	}
+
+	runnerCh, errCh := s.sched.GetRunner(c.Request.Context(), m, api.Options{}, nil)
+	var runner *runnerRef
+	select {
+	case runner = <-runnerCh:
+	case err := <-errCh:
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+
+	// Parse size (e.g., "1024x768") into width and height
+	width, height := int32(1024), int32(1024)
+	if req.Size != "" {
+		if _, err := fmt.Sscanf(req.Size, "%dx%d", &width, &height); err != nil {
+			c.JSON(http.StatusBadRequest, gin.H{"error": "invalid size format, expected WxH"})
+			return
+		}
+	}
+
+	var image []byte
+	err = runner.llama.Completion(c.Request.Context(), llm.CompletionRequest{
+		Prompt: req.Prompt,
+		Width:  width,
+		Height: height,
+	}, func(resp llm.CompletionResponse) {
+		if len(resp.Image) > 0 {
+			image = resp.Image
+		}
+	})
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+
+	c.JSON(http.StatusOK, gin.H{
+		"created": time.Now().Unix(),
+		"data":    []gin.H{{"b64_json": base64.StdEncoding.EncodeToString(image)}},
+	})
+}
+
 func (s *Server) ChatHandler(c *gin.Context) {
 	checkpointStart := time.Now()
 

server/sched_test.go

@@ -6,7 +6,6 @@ import (
 	"errors"
 	"log/slog"
 	"os"
-	"slices"
 	"testing"
 	"time"
 
@@ -17,7 +16,6 @@ import (
 	"github.com/ollama/ollama/fs/ggml"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/ml"
-	"github.com/ollama/ollama/types/model"
 )
 
 func TestMain(m *testing.M) {
@@ -807,32 +805,8 @@ func (s *mockLlm) GetDeviceInfos(ctx context.Context) []ml.DeviceInfo { return n
 func (s *mockLlm) HasExited() bool                                    { return false }
 func (s *mockLlm) GetActiveDeviceIDs() []ml.DeviceID                  { return nil }
 
-// TestImageGenCapabilityDetection verifies that models with "image" capability
-// are correctly identified and routed differently from language models.
-func TestImageGenCapabilityDetection(t *testing.T) {
-	// Model with image capability should be detected
-	imageModel := &Model{
-		Config: model.ConfigV2{
-			Capabilities: []string{"image"},
-		},
-	}
-	require.True(t, slices.Contains(imageModel.Config.Capabilities, "image"))
-
-	// Model without image capability should not be detected
-	langModel := &Model{
-		Config: model.ConfigV2{
-			Capabilities: []string{"completion"},
-		},
-	}
-	require.False(t, slices.Contains(langModel.Config.Capabilities, "image"))
-
-	// Empty capabilities should not match
-	emptyModel := &Model{}
-	require.False(t, slices.Contains(emptyModel.Config.Capabilities, "image"))
-}
-
 // TestImageGenRunnerCanBeEvicted verifies that an image generation model
-// loaded in the scheduler can be evicted by a language model request.
+// loaded in the scheduler can be evicted when idle.
 func TestImageGenRunnerCanBeEvicted(t *testing.T) {
 	ctx, done := context.WithTimeout(t.Context(), 500*time.Millisecond)
 	defer done()
@@ -864,3 +838,59 @@ func TestImageGenRunnerCanBeEvicted(t *testing.T) {
 	require.NotNil(t, runner)
 	require.Equal(t, "/fake/image/model", runner.modelPath)
 }
+
+// TestImageGenSchedulerCoexistence verifies that image generation models
+// can coexist with language models in the scheduler and VRAM is tracked correctly.
+func TestImageGenSchedulerCoexistence(t *testing.T) {
+	ctx, done := context.WithTimeout(t.Context(), 500*time.Millisecond)
+	defer done()
+
+	s := InitScheduler(ctx)
+	s.getGpuFn = getGpuFn
+	s.getSystemInfoFn = getSystemInfoFn
+
+	// Load both an imagegen runner and a language model runner
+	imageGenRunner := &runnerRef{
+		model:           &Model{Name: "flux", ModelPath: "/fake/flux/model"},
+		modelPath:       "/fake/flux/model",
+		llama:           &mockLlm{vramSize: 8 * format.GigaByte, vramByGPU: map[ml.DeviceID]uint64{{Library: "Metal"}: 8 * format.GigaByte}},
+		sessionDuration: 10 * time.Millisecond,
+		numParallel:     1,
+		refCount:        0,
+	}
+
+	langModelRunner := &runnerRef{
+		model:           &Model{Name: "llama3", ModelPath: "/fake/llama3/model"},
+		modelPath:       "/fake/llama3/model",
+		llama:           &mockLlm{vramSize: 4 * format.GigaByte, vramByGPU: map[ml.DeviceID]uint64{{Library: "Metal"}: 4 * format.GigaByte}},
+		sessionDuration: 10 * time.Millisecond,
+		numParallel:     1,
+		refCount:        0,
+	}
+
+	s.loadedMu.Lock()
+	s.loaded["/fake/flux/model"] = imageGenRunner
+	s.loaded["/fake/llama3/model"] = langModelRunner
+	s.loadedMu.Unlock()
+
+	// Verify both are loaded
+	s.loadedMu.Lock()
+	require.Len(t, s.loaded, 2)
+	require.NotNil(t, s.loaded["/fake/flux/model"])
+	require.NotNil(t, s.loaded["/fake/llama3/model"])
+	s.loadedMu.Unlock()
+
+	// Verify updateFreeSpace accounts for both
+	gpus := []ml.DeviceInfo{
+		{
+			DeviceID:    ml.DeviceID{Library: "Metal"},
+			TotalMemory: 24 * format.GigaByte,
+			FreeMemory:  24 * format.GigaByte,
+		},
+	}
+	s.updateFreeSpace(gpus)
+
+	// Free memory should be reduced by both models
+	expectedFree := uint64(24*format.GigaByte) - uint64(8*format.GigaByte) - uint64(4*format.GigaByte)
+	require.Equal(t, expectedFree, gpus[0].FreeMemory)
+}