refactor: let models declare their own default dimensions

Move width/height default application from server.go and runner.go
to the model code. Each model now handles its own defaults:
- zimage: uses imagegen.DefaultWidth/Height (1024x1024)
- qwen_image: uses imagegen.DefaultWidth/Height (1024x1024)
- qwen_image_edit: uses input image dimensions

The shared constants in defaults.go remain as fallbacks for models.

README.md

@@ -290,6 +290,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 
 ### Web & Desktop
 
+- [Onyx](https://github.com/onyx-dot-app/onyx)
 - [Open WebUI](https://github.com/open-webui/open-webui)
 - [SwiftChat (macOS with ReactNative)](https://github.com/aws-samples/swift-chat)
 - [Enchanted (macOS native)](https://github.com/AugustDev/enchanted)

api/types.go

@@ -127,6 +127,10 @@ type GenerateRequest struct {
 	// each with an associated log probability. Only applies when Logprobs is true.
 	// Valid values are 0-20. Default is 0 (only return the selected token's logprob).
 	TopLogprobs int `json:"top_logprobs,omitempty"`
+
+	// Size specifies the image dimensions for image generation models.
+	// Format: "WxH" (e.g., "1024x1024"). OpenAI-compatible.
+	Size string `json:"size,omitempty"`
 }
 
 // ChatRequest describes a request sent by [Client.Chat].

docs/docs.json

@@ -111,7 +111,9 @@
               "/integrations/zed",
               "/integrations/roo-code",
               "/integrations/n8n",
-              "/integrations/xcode"
+              "/integrations/xcode",
+              "/integrations/onyx",
+              "/integrations/marimo"
             ]
           },
           {

docs/integrations/marimo.mdx

@@ -0,0 +1,73 @@
+---
+title: marimo
+---
+
+## Install
+
+Install [marimo](https://marimo.io). You can use `pip` or `uv` for this. You 
+can also use `uv` to create a sandboxed environment for marimo by running:
+
+```
+uvx marimo edit --sandbox notebook.py
+```
+
+## Usage with Ollama
+
+1. In marimo, go to the user settings and go to the AI tab. From here
+you can find and configure Ollama as an AI provider. For local use you
+would typically point the base url to `http://localhost:11434/v1`.
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-settings.png" 
+    alt="Ollama settings in marimo"
+    width="50%"
+  />
+</div>
+
+2. Once the AI provider is set up, you can turn on/off specific AI models you'd like to access. 
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-models.png" 
+    alt="Selecting an Ollama model"
+    width="50%"
+  />
+</div>
+
+3. You can also add a model to the list of available models by scrolling to the bottom and using the UI there. 
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-add-model.png" 
+    alt="Adding a new Ollama model"
+    width="50%"
+  />
+</div>
+
+4. Once configured, you can now use Ollama for AI chats in marimo.
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-chat.png" 
+    alt="Configure code completion"
+    width="50%"
+  />
+</div>
+
+4. Alternatively, you can now use Ollama for **inline code completion** in marimo. This can be configured in the "AI Features" tab. 
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-code-completion.png" 
+    alt="Configure code completion"
+    width="50%"
+  />
+</div>
+
+
+## Connecting to ollama.com
+
+1. Sign in to ollama cloud via `ollama signin` 
+2. In the ollama model settings add a model that ollama hosts, like `gpt-oss:120b`.
+3. You can now refer to this model in marimo!

docs/integrations/onyx.mdx

@@ -0,0 +1,63 @@
+---
+title: Onyx
+---
+
+## Overview
+[Onyx](http://onyx.app/) is a self-hostable Chat UI that integrates with all Ollama models. Features include:
+- Creating custom Agents
+- Web search
+- Deep Research
+- RAG over uploaded documents and connected apps
+- Connectors to applications like Google Drive, Email, Slack, etc.
+- MCP and OpenAPI Actions support
+- Image generation
+- User/Groups management, RBAC, SSO, etc.
+
+Onyx can be deployed for single users or large organizations.
+
+## Install Onyx
+
+Deploy Onyx with the [quickstart guide](https://docs.onyx.app/deployment/getting_started/quickstart).
+
+<Info>
+Resourcing/scaling docs [here](https://docs.onyx.app/deployment/getting_started/resourcing).
+</Info>
+
+## Usage with Ollama 
+
+1. Login to your Onyx deployment (create an account first).
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/onyx-login.png" 
+    alt="Onyx Login Page"
+    width="75%"
+  />
+</div>
+2. In the set-up process select `Ollama` as the LLM provider.
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/onyx-ollama-llm.png" 
+    alt="Onyx Set Up Form"
+    width="75%"
+  />
+</div>
+3. Provide your **Ollama API URL** and select your models.
+<Note>If you're running Onyx in Docker, to access your computer's local network use `http://host.docker.internal` instead of `http://127.0.0.1`.</Note>
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/onyx-ollama-form.png" 
+    alt="Selecting Ollama Models"
+    width="75%"
+  />
+</div>
+
+You can also easily connect up Onyx Cloud with the `Ollama Cloud` tab of the setup.
+
+## Send your first query
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/onyx-query.png" 
+    alt="Onyx Query Example"
+    width="75%"
+  />
+</div>

llm/server.go

@@ -1464,6 +1464,10 @@ type CompletionRequest struct {
 
 	// TopLogprobs specifies the number of most likely alternative tokens to return (0-20)
 	TopLogprobs int
+
+	// Size specifies image dimensions for image generation models.
+	// Format: "WxH" (e.g., "1024x1024"). OpenAI-compatible.
+	Size string
 }
 
 // DoneReason represents the reason why a completion response is done

server/routes.go

@@ -216,7 +216,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 
 	// 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)
+		imagegenapi.HandleGenerateRequest(c, s, &req, streamResponse)
 		return
 	}
 

server/sched.go

@@ -574,7 +574,6 @@ func (s *Scheduler) loadImageGen(req *LlmRequest) bool {
 		Options:         &req.opts,
 		loading:         false,
 		sessionDuration: sessionDuration,
-		refCount:        1,
 	}
 
 	s.loadedMu.Lock()

x/imagegen/api/handler.go

@@ -2,8 +2,8 @@ package api
 
 import (
 	"fmt"
+	"log/slog"
 	"net/http"
-	"strconv"
 	"strings"
 	"time"
 
@@ -50,7 +50,7 @@ func ImageGenerationHandler(c *gin.Context, scheduler RunnerScheduler) {
 		req.N = 1
 	}
 	if req.Size == "" {
-		req.Size = "1024x1024"
+		req.Size = fmt.Sprintf("%dx%d", imagegen.DefaultWidth, imagegen.DefaultHeight)
 	}
 	if req.ResponseFormat == "" {
 		req.ResponseFormat = "b64_json"
@@ -62,16 +62,8 @@ func ImageGenerationHandler(c *gin.Context, scheduler RunnerScheduler) {
 		return
 	}
 
-	// Parse size
-	width, height := parseSize(req.Size)
-
-	// Build options - we repurpose NumCtx/NumGPU for width/height
-	opts := api.Options{}
-	opts.NumCtx = int(width)
-	opts.NumGPU = int(height)
-
 	// Schedule runner
-	runner, err := scheduler.ScheduleImageGenRunner(c, req.Model, opts, nil)
+	runner, err := scheduler.ScheduleImageGenRunner(c, req.Model, api.Options{}, nil)
 	if err != nil {
 		status := http.StatusInternalServerError
 		if strings.Contains(err.Error(), "not found") {
@@ -81,10 +73,10 @@ func ImageGenerationHandler(c *gin.Context, scheduler RunnerScheduler) {
 		return
 	}
 
-	// Build completion request
+	// Build completion request with size (OpenAI format)
 	completionReq := llm.CompletionRequest{
-		Prompt:  req.Prompt,
-		Options: &opts,
+		Prompt: req.Prompt,
+		Size:   req.Size,
 	}
 
 	if req.Stream {
@@ -134,22 +126,6 @@ func handleNonStreamingResponse(c *gin.Context, runner llm.LlamaServer, req llm.
 	c.JSON(http.StatusOK, buildResponse(imageBase64, format))
 }
 
-func parseSize(size string) (int32, int32) {
-	parts := strings.Split(size, "x")
-	if len(parts) != 2 {
-		return 1024, 1024
-	}
-	w, _ := strconv.Atoi(parts[0])
-	h, _ := strconv.Atoi(parts[1])
-	if w == 0 {
-		w = 1024
-	}
-	if h == 0 {
-		h = 1024
-	}
-	return int32(w), int32(h)
-}
-
 func extractBase64(content string) string {
 	if strings.HasPrefix(content, "IMAGE_BASE64:") {
 		return content[13:]
@@ -185,20 +161,18 @@ func buildResponse(imageBase64, format string) ImageGenerationResponse {
 
 // HandleGenerateRequest handles Ollama /api/generate requests for image gen models.
 // This allows routes.go to delegate image generation with minimal code.
-func HandleGenerateRequest(c *gin.Context, scheduler RunnerScheduler, modelName, prompt string, keepAlive *api.Duration, streamFn func(c *gin.Context, ch chan any)) {
-	opts := api.Options{}
-
+func HandleGenerateRequest(c *gin.Context, scheduler RunnerScheduler, req *api.GenerateRequest, streamFn func(c *gin.Context, ch chan any)) {
 	// Schedule runner
-	runner, err := scheduler.ScheduleImageGenRunner(c, modelName, opts, keepAlive)
+	runner, err := scheduler.ScheduleImageGenRunner(c, req.Model, api.Options{}, req.KeepAlive)
 	if err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
 	}
 
-	// Build completion request
+	// Build completion request with size (OpenAI format)
 	completionReq := llm.CompletionRequest{
-		Prompt:  prompt,
-		Options: &opts,
+		Prompt: req.Prompt,
+		Size:   req.Size,
 	}
 
 	// Stream responses via channel
@@ -207,15 +181,14 @@ func HandleGenerateRequest(c *gin.Context, scheduler RunnerScheduler, modelName,
 		defer close(ch)
 		err := runner.Completion(c.Request.Context(), completionReq, func(resp llm.CompletionResponse) {
 			ch <- GenerateResponse{
-				Model:     modelName,
+				Model:     req.Model,
 				CreatedAt: time.Now().UTC(),
 				Response:  resp.Content,
 				Done:      resp.Done,
 			}
 		})
 		if err != nil {
-			// Log error but don't block - channel is already being consumed
-			_ = err
+			slog.Error("image generation failed", "model", req.Model, "error", err)
 		}
 	}()
 

x/imagegen/cli.go

@@ -37,9 +37,9 @@ type ImageGenOptions struct {
 // DefaultOptions returns the default image generation options.
 func DefaultOptions() ImageGenOptions {
 	return ImageGenOptions{
-		Width:  1024,
-		Height: 1024,
-		Steps:  9,
+		Width:  DefaultWidth,
+		Height: DefaultHeight,
+		Steps:  0, // 0 means model default
 		Seed:   0, // 0 means random
 	}
 }
@@ -107,9 +107,9 @@ func GetModelInfo(modelName string) (*ModelInfo, error) {
 // RegisterFlags adds image generation flags to the given command.
 // Flags are hidden since they only apply to image generation models.
 func RegisterFlags(cmd *cobra.Command) {
-	cmd.Flags().Int("width", 1024, "Image width")
-	cmd.Flags().Int("height", 1024, "Image height")
-	cmd.Flags().Int("steps", 9, "Denoising steps")
+	cmd.Flags().Int("width", DefaultWidth, "Image width")
+	cmd.Flags().Int("height", DefaultHeight, "Image height")
+	cmd.Flags().Int("steps", 0, "Denoising steps (0 = model default)")
 	cmd.Flags().Int("seed", 0, "Random seed (0 for random)")
 	cmd.Flags().String("negative", "", "Negative prompt")
 	cmd.Flags().MarkHidden("width")
@@ -158,17 +158,10 @@ func generateImageWithOptions(cmd *cobra.Command, modelName, prompt string, keep
 		return err
 	}
 
-	// Build request with image gen options encoded in Options fields
-	// NumCtx=width, NumGPU=height, NumPredict=steps, Seed=seed
 	req := &api.GenerateRequest{
 		Model:  modelName,
 		Prompt: prompt,
-		Options: map[string]any{
-			"num_ctx":     opts.Width,
-			"num_gpu":     opts.Height,
-			"num_predict": opts.Steps,
-			"seed":        opts.Seed,
-		},
+		Size:   fmt.Sprintf("%dx%d", opts.Width, opts.Height),
 	}
 	if keepAlive != nil {
 		req.KeepAlive = keepAlive

x/imagegen/cmd/engine/main.go

@@ -12,6 +12,7 @@ import (
 	"path/filepath"
 	"runtime/pprof"
 
+	"github.com/ollama/ollama/x/imagegen"
 	"github.com/ollama/ollama/x/imagegen/mlx"
 	"github.com/ollama/ollama/x/imagegen/models/gemma3"
 	"github.com/ollama/ollama/x/imagegen/models/gpt_oss"
@@ -46,9 +47,9 @@ func main() {
 	imagePath := flag.String("image", "", "Image path for multimodal models")
 
 	// Image generation params
-	width := flag.Int("width", 1024, "Image width")
-	height := flag.Int("height", 1024, "Image height")
-	steps := flag.Int("steps", 9, "Denoising steps")
+	width := flag.Int("width", imagegen.DefaultWidth, "Image width")
+	height := flag.Int("height", imagegen.DefaultHeight, "Image height")
+	steps := flag.Int("steps", 0, "Denoising steps (0 = model default)")
 	seed := flag.Int64("seed", 42, "Random seed")
 	out := flag.String("output", "output.png", "Output path")
 
@@ -149,10 +150,10 @@ func main() {
 		// unless explicitly overridden from defaults
 		editWidth := int32(0)
 		editHeight := int32(0)
-		if *width != 1024 {
+		if *width != imagegen.DefaultWidth {
 			editWidth = int32(*width)
 		}
-		if *height != 1024 {
+		if *height != imagegen.DefaultHeight {
 			editHeight = int32(*height)
 		}
 

x/imagegen/defaults.go

@@ -0,0 +1,7 @@
+package imagegen
+
+// Default image generation parameters.
+const (
+	DefaultWidth  = 1024
+	DefaultHeight = 1024
+)

x/imagegen/memory.go

@@ -95,8 +95,3 @@ func EstimateVRAM(modelName string) uint64 {
 	}
 	return 21 * GB
 }
-
-// HasTensorLayers checks if the given model has tensor layers.
-func HasTensorLayers(modelName string) bool {
-	return ResolveModelName(modelName) != ""
-}

x/imagegen/memory_test.go

@@ -94,13 +94,6 @@ func TestEstimateVRAMDefault(t *testing.T) {
 	}
 }
 
-func TestHasTensorLayers(t *testing.T) {
-	// Non-existent model should return false
-	if HasTensorLayers("nonexistent-model") {
-		t.Error("HasTensorLayers() should return false for non-existent model")
-	}
-}
-
 func TestResolveModelName(t *testing.T) {
 	// Non-existent model should return empty string
 	result := ResolveModelName("nonexistent-model")

x/imagegen/models/qwen_image/qwen_image.go

@@ -9,6 +9,7 @@ import (
 	"path/filepath"
 	"time"
 
+	"github.com/ollama/ollama/x/imagegen"
 	"github.com/ollama/ollama/x/imagegen/cache"
 	"github.com/ollama/ollama/x/imagegen/mlx"
 	"github.com/ollama/ollama/x/imagegen/tokenizer"
@@ -166,13 +167,13 @@ func (m *Model) GenerateImage(ctx context.Context, prompt string, width, height
 func (m *Model) generate(cfg *GenerateConfig) (*mlx.Array, error) {
 	// Apply defaults
 	if cfg.Width <= 0 {
-		cfg.Width = 1024
+		cfg.Width = imagegen.DefaultWidth
 	}
 	if cfg.Height <= 0 {
-		cfg.Height = 1024
+		cfg.Height = imagegen.DefaultHeight
 	}
 	if cfg.Steps <= 0 {
-		cfg.Steps = 30
+		cfg.Steps = 50
 	}
 	if cfg.CFGScale <= 0 {
 		cfg.CFGScale = 4.0

x/imagegen/models/zimage/zimage.go

@@ -188,13 +188,13 @@ func (m *Model) GenerateImage(ctx context.Context, prompt string, width, height
 func (m *Model) generate(ctx context.Context, cfg *GenerateConfig) (*mlx.Array, error) {
 	// Apply defaults
 	if cfg.Width <= 0 {
-		cfg.Width = 1024
+		cfg.Width = imagegen.DefaultWidth
 	}
 	if cfg.Height <= 0 {
-		cfg.Height = 1024
+		cfg.Height = imagegen.DefaultHeight
 	}
 	if cfg.Steps <= 0 {
-		cfg.Steps = 9 // Turbo default
+		cfg.Steps = 9 // Z-Image turbo default
 	}
 	if cfg.CFGScale <= 0 {
 		cfg.CFGScale = 4.0

x/imagegen/runner/runner.go

@@ -136,16 +136,8 @@ func (s *Server) completionHandler(w http.ResponseWriter, r *http.Request) {
 	s.mu.Lock()
 	defer s.mu.Unlock()
 
-	// Apply defaults
-	if req.Width <= 0 {
-		req.Width = 1024
-	}
-	if req.Height <= 0 {
-		req.Height = 1024
-	}
-	if req.Steps <= 0 {
-		req.Steps = 9
-	}
+	// Model applies its own defaults for width/height/steps
+	// Only seed needs to be set here if not provided
 	if req.Seed <= 0 {
 		req.Seed = time.Now().UnixNano()
 	}

x/imagegen/server.go

@@ -33,10 +33,12 @@ type Server struct {
 	vramSize    uint64
 	done        chan error
 	client      *http.Client
-	lastErr     string // Last stderr line for error reporting
-	lastErrLock sync.Mutex
+	stderrLines []string // Recent stderr lines for error reporting (max 10)
+	stderrLock  sync.Mutex
 }
 
+const maxStderrLines = 10
+
 // completionRequest is sent to the subprocess
 type completionRequest struct {
 	Prompt string `json:"prompt"`
@@ -139,10 +141,13 @@ func NewServer(modelName string) (*Server, error) {
 		for scanner.Scan() {
 			line := scanner.Text()
 			slog.Warn("image-runner", "msg", line)
-			// Capture last error line for better error reporting
-			s.lastErrLock.Lock()
-			s.lastErr = line
-			s.lastErrLock.Unlock()
+			// Capture recent stderr lines for error reporting
+			s.stderrLock.Lock()
+			s.stderrLines = append(s.stderrLines, line)
+			if len(s.stderrLines) > maxStderrLines {
+				s.stderrLines = s.stderrLines[1:]
+			}
+			s.stderrLock.Unlock()
 		}
 	}()
 
@@ -171,7 +176,9 @@ func (s *Server) ModelPath() string {
 	return s.modelName
 }
 
-// Load is called by the scheduler after the server is created.
+// Load is a no-op for image generation models.
+// Unlike LLM models, imagegen models are loaded by the subprocess at startup
+// rather than through this interface method.
 func (s *Server) Load(ctx context.Context, systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, requireFull bool) ([]ml.DeviceID, error) {
 	return nil, nil
 }
@@ -204,20 +211,16 @@ func (s *Server) waitUntilRunning() error {
 	for {
 		select {
 		case err := <-s.done:
-			// Include last stderr line for better error context
-			s.lastErrLock.Lock()
-			lastErr := s.lastErr
-			s.lastErrLock.Unlock()
-			if lastErr != "" {
-				return fmt.Errorf("image runner failed: %s (exit: %v)", lastErr, err)
+			// Include recent stderr lines for better error context
+			stderrContext := s.getStderrContext()
+			if stderrContext != "" {
+				return fmt.Errorf("image runner failed: %s (exit: %v)", stderrContext, err)
 			}
 			return fmt.Errorf("image runner exited unexpectedly: %w", err)
 		case <-timeout:
-			s.lastErrLock.Lock()
-			lastErr := s.lastErr
-			s.lastErrLock.Unlock()
-			if lastErr != "" {
-				return fmt.Errorf("timeout waiting for image runner: %s", lastErr)
+			stderrContext := s.getStderrContext()
+			if stderrContext != "" {
+				return fmt.Errorf("timeout waiting for image runner: %s", stderrContext)
 			}
 			return errors.New("timeout waiting for image runner to start")
 		case <-ticker.C:
@@ -229,34 +232,36 @@ func (s *Server) waitUntilRunning() error {
 	}
 }
 
-// WaitUntilRunning implements the LlamaServer interface (no-op since NewServer waits).
+// getStderrContext returns recent stderr lines joined as a single string.
+func (s *Server) getStderrContext() string {
+	s.stderrLock.Lock()
+	defer s.stderrLock.Unlock()
+	if len(s.stderrLines) == 0 {
+		return ""
+	}
+	return strings.Join(s.stderrLines, "; ")
+}
+
+// WaitUntilRunning is a no-op for image generation models.
+// NewServer already blocks until the subprocess is ready, so this method
+// returns immediately. Required by the llm.LlamaServer interface.
 func (s *Server) WaitUntilRunning(ctx context.Context) error {
 	return nil
 }
 
 // Completion generates an image from the prompt via the subprocess.
 func (s *Server) Completion(ctx context.Context, req llm.CompletionRequest, fn func(llm.CompletionResponse)) error {
-	// Build request
+	// Build request - let the model apply its own defaults for unspecified values
 	creq := completionRequest{
 		Prompt: req.Prompt,
-		Width:  1024,
-		Height: 1024,
-		Steps:  9,
 		Seed:   time.Now().UnixNano(),
 	}
 
-	if req.Options != nil {
-		if req.Options.NumCtx > 0 && req.Options.NumCtx <= 4096 {
-			creq.Width = int32(req.Options.NumCtx)
-		}
-		if req.Options.NumGPU > 0 && req.Options.NumGPU <= 4096 {
-			creq.Height = int32(req.Options.NumGPU)
-		}
-		if req.Options.NumPredict > 0 && req.Options.NumPredict <= 100 {
-			creq.Steps = req.Options.NumPredict
-		}
-		if req.Options.Seed > 0 {
-			creq.Seed = int64(req.Options.Seed)
+	// Parse size string (OpenAI format: "WxH") - only set if provided
+	if req.Size != "" {
+		if w, h := parseSize(req.Size); w > 0 && h > 0 {
+			creq.Width = w
+			creq.Height = h
 		}
 	}
 
@@ -346,17 +351,20 @@ func (s *Server) VRAMByGPU(id ml.DeviceID) uint64 {
 	return s.vramSize
 }
 
-// Embedding is not supported for image generation models.
+// Embedding returns an error as image generation models don't produce embeddings.
+// Required by the llm.LlamaServer interface.
 func (s *Server) Embedding(ctx context.Context, input string) ([]float32, int, error) {
 	return nil, 0, errors.New("embedding not supported for image generation models")
 }
 
-// Tokenize is not supported for image generation models.
+// Tokenize returns an error as image generation uses internal tokenization.
+// Required by the llm.LlamaServer interface.
 func (s *Server) Tokenize(ctx context.Context, content string) ([]int, error) {
 	return nil, errors.New("tokenize not supported for image generation models")
 }
 
-// Detokenize is not supported for image generation models.
+// Detokenize returns an error as image generation uses internal tokenization.
+// Required by the llm.LlamaServer interface.
 func (s *Server) Detokenize(ctx context.Context, tokens []int) (string, error) {
 	return "", errors.New("detokenize not supported for image generation models")
 }
@@ -376,7 +384,8 @@ func (s *Server) GetPort() int {
 	return s.port
 }
 
-// GetDeviceInfos returns nil since we don't track GPU info.
+// GetDeviceInfos returns nil as GPU tracking is handled by the subprocess.
+// Required by the llm.LlamaServer interface.
 func (s *Server) GetDeviceInfos(ctx context.Context) []ml.DeviceInfo {
 	return nil
 }
@@ -393,3 +402,14 @@ func (s *Server) HasExited() bool {
 
 // Ensure Server implements llm.LlamaServer
 var _ llm.LlamaServer = (*Server)(nil)
+
+// parseSize parses an OpenAI-style size string "WxH" into width and height.
+func parseSize(size string) (int32, int32) {
+	parts := strings.Split(size, "x")
+	if len(parts) != 2 {
+		return 0, 0
+	}
+	w, _ := strconv.Atoi(parts[0])
+	h, _ := strconv.Atoi(parts[1])
+	return int32(w), int32(h)
+}