create: read draft architecture from its config.json

Previously the draft architecture was hardcoded to
Gemma4AssistantForCausalLM. Read it from the draft model's config so
any draft architecture can be packaged.

x/create/client/create.go

@@ -243,6 +243,40 @@ func appendLayersManifestWriter(next create.ManifestWriter, extra []create.Layer
 	}
 }
 
+func draftMetadata(draftDir string) (*model.Draft, error) {
+	configPath := filepath.Join(draftDir, "config.json")
+	data, err := os.ReadFile(configPath)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read draft config %s: %w", configPath, err)
+	}
+
+	var cfg struct {
+		Architectures []string `json:"architectures"`
+		ModelType     string   `json:"model_type"`
+	}
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return nil, fmt.Errorf("failed to parse draft config %s: %w", configPath, err)
+	}
+
+	arch := ""
+	if len(cfg.Architectures) > 0 {
+		arch = cfg.Architectures[0]
+	}
+	if arch == "" {
+		arch = cfg.ModelType
+	}
+	if arch == "" {
+		return nil, fmt.Errorf("draft architecture not found in %s", configPath)
+	}
+
+	return &model.Draft{
+		ModelFormat:  "safetensors",
+		Architecture: arch,
+		TensorPrefix: "draft.",
+		Config:       "draft/config.json",
+	}, nil
+}
+
 func createModelFromBaseWithDraft(opts CreateOptions, draftLayers []create.LayerInfo, progressFn func(string)) error {
 	progressFn(fmt.Sprintf("loading base model %s", opts.ModelDir))
 	baseManifest, err := imagemanifest.LoadManifest(opts.ModelDir)
@@ -487,12 +521,11 @@ func newManifestWriter(opts CreateOptions, capabilities []string, parserName, re
 		configData.Parser = resolveParserName(opts.Modelfile, parserName)
 		configData.Renderer = resolveRendererName(opts.Modelfile, rendererName)
 		if opts.Modelfile != nil && opts.Modelfile.Draft != "" {
-			configData.Draft = &model.Draft{
-				ModelFormat:  "safetensors",
-				Architecture: "Gemma4AssistantForCausalLM",
-				TensorPrefix: "draft.",
-				Config:       "draft/config.json",
+			draft, err := draftMetadata(opts.Modelfile.Draft)
+			if err != nil {
+				return err
 			}
+			configData.Draft = draft
 		}
 		configJSON, err := json.Marshal(configData)
 		if err != nil {

x/create/client/create_test.go

@@ -544,10 +544,15 @@ func TestNewManifestWriter_PopulatesFileTypeFromQuantize(t *testing.T) {
 func TestNewManifestWriter_PopulatesDraftMetadata(t *testing.T) {
 	t.Setenv("OLLAMA_MODELS", t.TempDir())
 
+	draftDir := t.TempDir()
+	if err := os.WriteFile(filepath.Join(draftDir, "config.json"), []byte(`{"architectures":["DFlashDraftModel"],"model_type":"qwen3"}`), 0o644); err != nil {
+		t.Fatalf("WriteFile() error = %v", err)
+	}
+
 	opts := CreateOptions{
 		ModelName: "test-draft",
 		ModelDir:  t.TempDir(),
-		Modelfile: &ModelfileConfig{Draft: "/tmp/assistant"},
+		Modelfile: &ModelfileConfig{Draft: draftDir},
 	}
 
 	writer := newManifestWriter(opts, []string{"completion"}, "gemma4", "gemma4")
@@ -581,6 +586,9 @@ func TestNewManifestWriter_PopulatesDraftMetadata(t *testing.T) {
 	if cfg.Draft.TensorPrefix != "draft." || cfg.Draft.Config != "draft/config.json" {
 		t.Fatalf("Draft = %#v, want draft prefix/config", cfg.Draft)
 	}
+	if cfg.Draft.Architecture != "DFlashDraftModel" {
+		t.Fatalf("Draft architecture = %q, want DFlashDraftModel", cfg.Draft.Architecture)
+	}
 }
 
 func TestSupportsThinking(t *testing.T) {