x models: add glm 4.7 flash to mlx engine

manifest/layer.go

@@ -1,4 +1,4 @@
-package server
+package manifest
 
 import (
 	"crypto/sha256"
@@ -14,7 +14,7 @@ type Layer struct {
 	Size      int64  `json:"size"`
 	From      string `json:"from,omitempty"`
 	Name      string `json:"name,omitempty"` // tensor name, e.g., "text_encoder/model.embed_tokens.weight"
-	status    string
+	Status    string `json:"-"`
 }
 
 const (
@@ -22,7 +22,7 @@ const (
 )
 
 func NewLayer(r io.Reader, mediatype string) (Layer, error) {
-	blobs, err := GetBlobsPath("")
+	blobs, err := BlobsPath("")
 	if err != nil {
 		return Layer{}, err
 	}
@@ -45,7 +45,7 @@ func NewLayer(r io.Reader, mediatype string) (Layer, error) {
 	}
 
 	digest := fmt.Sprintf("sha256:%x", sha256sum.Sum(nil))
-	blob, err := GetBlobsPath(digest)
+	blob, err := BlobsPath(digest)
 	if err != nil {
 		return Layer{}, err
 	}
@@ -65,7 +65,7 @@ func NewLayer(r io.Reader, mediatype string) (Layer, error) {
 		MediaType: mediatype,
 		Digest:    digest,
 		Size:      n,
-		status:    fmt.Sprintf("%s %s", status, digest),
+		Status:    fmt.Sprintf("%s %s", status, digest),
 	}, nil
 }
 
@@ -74,7 +74,7 @@ func NewLayerFromLayer(digest, mediatype, from string) (Layer, error) {
 		return Layer{}, errors.New("creating new layer from layer with empty digest")
 	}
 
-	blob, err := GetBlobsPath(digest)
+	blob, err := BlobsPath(digest)
 	if err != nil {
 		return Layer{}, err
 	}
@@ -89,7 +89,7 @@ func NewLayerFromLayer(digest, mediatype, from string) (Layer, error) {
 		Digest:    digest,
 		Size:      fi.Size(),
 		From:      from,
-		status:    fmt.Sprintf("using existing layer %s", digest),
+		Status:    fmt.Sprintf("using existing layer %s", digest),
 	}, nil
 }
 
@@ -98,7 +98,7 @@ func (l *Layer) Open() (io.ReadSeekCloser, error) {
 		return nil, errors.New("opening layer with empty digest")
 	}
 
-	blob, err := GetBlobsPath(l.Digest)
+	blob, err := BlobsPath(l.Digest)
 	if err != nil {
 		return nil, err
 	}
@@ -126,7 +126,7 @@ func (l *Layer) Remove() error {
 		}
 	}
 
-	blob, err := GetBlobsPath(l.Digest)
+	blob, err := BlobsPath(l.Digest)
 	if err != nil {
 		return err
 	}

manifest/manifest.go

@@ -1,10 +1,9 @@
-package server
+package manifest
 
 import (
 	"crypto/sha256"
 	"encoding/hex"
 	"encoding/json"
-	"errors"
 	"fmt"
 	"io"
 	"log/slog"
@@ -33,12 +32,38 @@ func (m *Manifest) Size() (size int64) {
 	return
 }
 
+func (m *Manifest) Digest() string {
+	return m.digest
+}
+
+func (m *Manifest) FileInfo() os.FileInfo {
+	return m.fi
+}
+
+// ReadConfigJSON reads and unmarshals a config layer as JSON.
+func (m *Manifest) ReadConfigJSON(configPath string, v any) error {
+	for _, layer := range m.Layers {
+		if layer.MediaType == "application/vnd.ollama.image.json" && layer.Name == configPath {
+			blobPath, err := BlobsPath(layer.Digest)
+			if err != nil {
+				return err
+			}
+			data, err := os.ReadFile(blobPath)
+			if err != nil {
+				return err
+			}
+			return json.Unmarshal(data, v)
+		}
+	}
+	return fmt.Errorf("config %q not found in manifest", configPath)
+}
+
 func (m *Manifest) Remove() error {
 	if err := os.Remove(m.filepath); err != nil {
 		return err
 	}
 
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return err
 	}
@@ -70,11 +95,11 @@ func (m *Manifest) RemoveLayers() error {
 		if _, used := inUse[layer.Digest]; used {
 			continue
 		}
-		blob, err := GetBlobsPath(layer.Digest)
+		blob, err := BlobsPath(layer.Digest)
 		if err != nil {
 			return err
 		}
-		if err := os.Remove(blob); errors.Is(err, os.ErrNotExist) {
+		if err := os.Remove(blob); os.IsNotExist(err) {
 			slog.Debug("layer does not exist", "digest", layer.Digest)
 		} else if err != nil {
 			return err
@@ -89,7 +114,7 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) {
 		return nil, model.Unqualified(n)
 	}
 
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return nil, err
 	}
@@ -121,7 +146,7 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) {
 }
 
 func WriteManifest(name model.Name, config Layer, layers []Layer) error {
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return err
 	}
@@ -148,7 +173,7 @@ func WriteManifest(name model.Name, config Layer, layers []Layer) error {
 }
 
 func Manifests(continueOnError bool) (map[model.Name]*Manifest, error) {
-	manifests, err := GetManifestPath()
+	manifests, err := Path()
 	if err != nil {
 		return nil, err
 	}

manifest/manifest_test.go

@@ -1,4 +1,4 @@
-package server
+package manifest
 
 import (
 	"encoding/json"

manifest/paths.go

@@ -0,0 +1,95 @@
+package manifest
+
+import (
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strings"
+
+	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/types/model"
+)
+
+var ErrInvalidDigestFormat = errors.New("invalid digest format")
+
+func Path() (string, error) {
+	path := filepath.Join(envconfig.Models(), "manifests")
+	if err := os.MkdirAll(path, 0o755); err != nil {
+		return "", fmt.Errorf("%w: ensure path elements are traversable", err)
+	}
+
+	return path, nil
+}
+
+// PathForName returns the path to the manifest file for a specific model name.
+func PathForName(n model.Name) (string, error) {
+	if !n.IsValid() {
+		return "", os.ErrNotExist
+	}
+
+	manifests, err := Path()
+	if err != nil {
+		return "", err
+	}
+
+	return filepath.Join(manifests, n.Filepath()), nil
+}
+
+func BlobsPath(digest string) (string, error) {
+	// only accept actual sha256 digests
+	pattern := "^sha256[:-][0-9a-fA-F]{64}$"
+	re := regexp.MustCompile(pattern)
+
+	if digest != "" && !re.MatchString(digest) {
+		return "", ErrInvalidDigestFormat
+	}
+
+	digest = strings.ReplaceAll(digest, ":", "-")
+	path := filepath.Join(envconfig.Models(), "blobs", digest)
+	dirPath := filepath.Dir(path)
+	if digest == "" {
+		dirPath = path
+	}
+
+	if err := os.MkdirAll(dirPath, 0o755); err != nil {
+		return "", fmt.Errorf("%w: ensure path elements are traversable", err)
+	}
+
+	return path, nil
+}
+
+// PruneDirectory removes empty directories recursively.
+func PruneDirectory(path string) error {
+	info, err := os.Lstat(path)
+	if err != nil {
+		return err
+	}
+
+	if info.IsDir() && info.Mode()&os.ModeSymlink == 0 {
+		entries, err := os.ReadDir(path)
+		if err != nil {
+			return err
+		}
+
+		for _, entry := range entries {
+			if err := PruneDirectory(filepath.Join(path, entry.Name())); err != nil {
+				return err
+			}
+		}
+
+		entries, err = os.ReadDir(path)
+		if err != nil {
+			return err
+		}
+
+		if len(entries) > 0 {
+			return nil
+		}
+
+		return os.Remove(path)
+	}
+
+	return nil
+}

server/create.go

@@ -28,6 +28,7 @@ import (
 	"github.com/ollama/ollama/format"
 	ofs "github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/manifest"
 	"github.com/ollama/ollama/template"
 	"github.com/ollama/ollama/types/errtypes"
 	"github.com/ollama/ollama/types/model"
@@ -90,7 +91,7 @@ func (s *Server) CreateHandler(c *gin.Context) {
 			ch <- resp
 		}
 
-		oldManifest, _ := ParseNamedManifest(name)
+		oldManifest, _ := manifest.ParseNamedManifest(name)
 
 		var baseLayers []*layerGGML
 		var err error
@@ -123,9 +124,9 @@ func (s *Server) CreateHandler(c *gin.Context) {
 				}
 
 				if err == nil && !remote && (config.Renderer == "" || config.Parser == "" || config.Requires == "") {
-					manifest, mErr := ParseNamedManifest(fromName)
-					if mErr == nil && manifest.Config.Digest != "" {
-						configPath, pErr := GetBlobsPath(manifest.Config.Digest)
+					mf, mErr := manifest.ParseNamedManifest(fromName)
+					if mErr == nil && mf.Config.Digest != "" {
+						configPath, pErr := manifest.BlobsPath(mf.Config.Digest)
 						if pErr == nil {
 							if cfgFile, fErr := os.Open(configPath); fErr == nil {
 								var baseConfig model.ConfigV2
@@ -342,7 +343,7 @@ func detectModelTypeFromFiles(files map[string]string) string {
 			return "gguf"
 		} else {
 			// try to see if we can find a gguf file even without the file extension
-			blobPath, err := GetBlobsPath(files[fn])
+			blobPath, err := manifest.BlobsPath(files[fn])
 			if err != nil {
 				slog.Error("error getting blobs path", "file", fn)
 				return ""
@@ -394,7 +395,7 @@ func convertFromSafetensors(files map[string]string, baseLayers []*layerGGML, is
 			return nil, fmt.Errorf("%w: %s: %s", errFilePath, err, fp)
 		}
 
-		blobPath, err := GetBlobsPath(digest)
+		blobPath, err := manifest.BlobsPath(digest)
 		if err != nil {
 			return nil, err
 		}
@@ -432,7 +433,7 @@ func convertFromSafetensors(files map[string]string, baseLayers []*layerGGML, is
 		return nil, err
 	}
 
-	layer, err := NewLayer(t, mediaType)
+	layer, err := manifest.NewLayer(t, mediaType)
 	if err != nil {
 		return nil, err
 	}
@@ -465,7 +466,7 @@ func kvFromLayers(baseLayers []*layerGGML) (ofs.Config, error) {
 }
 
 func createModel(r api.CreateRequest, name model.Name, baseLayers []*layerGGML, config *model.ConfigV2, fn func(resp api.ProgressResponse)) (err error) {
-	var layers []Layer
+	var layers []manifest.Layer
 	for _, layer := range baseLayers {
 		if layer.GGML != nil {
 			quantType := strings.ToUpper(cmp.Or(r.Quantize, r.Quantization))
@@ -550,13 +551,13 @@ func createModel(r api.CreateRequest, name model.Name, baseLayers []*layerGGML,
 	}
 
 	for _, layer := range layers {
-		if layer.status != "" {
-			fn(api.ProgressResponse{Status: layer.status})
+		if layer.Status != "" {
+			fn(api.ProgressResponse{Status: layer.Status})
 		}
 	}
 
 	fn(api.ProgressResponse{Status: "writing manifest"})
-	if err := WriteManifest(name, *configLayer, layers); err != nil {
+	if err := manifest.WriteManifest(name, *configLayer, layers); err != nil {
 		return err
 	}
 
@@ -577,7 +578,7 @@ func quantizeLayer(layer *layerGGML, quantizeType string, fn func(resp api.Progr
 		return nil, err
 	}
 
-	blob, err := GetBlobsPath(layer.Digest)
+	blob, err := manifest.BlobsPath(layer.Digest)
 	if err != nil {
 		return nil, err
 	}
@@ -599,7 +600,7 @@ func quantizeLayer(layer *layerGGML, quantizeType string, fn func(resp api.Progr
 	}
 	temp.Seek(0, io.SeekStart)
 	fn(api.ProgressResponse{Status: "verifying conversion"})
-	newLayer, err := NewLayer(temp, layer.MediaType)
+	newLayer, err := manifest.NewLayer(temp, layer.MediaType)
 	if err != nil {
 		return nil, err
 	}
@@ -619,7 +620,7 @@ func ggufLayers(digest string, fn func(resp api.ProgressResponse)) ([]*layerGGML
 	var layers []*layerGGML
 
 	fn(api.ProgressResponse{Status: "parsing GGUF"})
-	blobPath, err := GetBlobsPath(digest)
+	blobPath, err := manifest.BlobsPath(digest)
 	if err != nil {
 		return nil, err
 	}
@@ -654,7 +655,7 @@ func ggufLayers(digest string, fn func(resp api.ProgressResponse)) ([]*layerGGML
 		mediatype = "application/vnd.ollama.image.projector"
 	}
 
-	layer, err := NewLayerFromLayer(digest, mediatype, blob.Name())
+	layer, err := manifest.NewLayerFromLayer(digest, mediatype, blob.Name())
 	if err != nil {
 		slog.Debug("could not create new layer from layer", "error", err)
 		return nil, err
@@ -665,8 +666,8 @@ func ggufLayers(digest string, fn func(resp api.ProgressResponse)) ([]*layerGGML
 	return detectChatTemplate(layers)
 }
 
-func removeLayer(layers []Layer, mediatype string) []Layer {
-	return slices.DeleteFunc(layers, func(layer Layer) bool {
+func removeLayer(layers []manifest.Layer, mediatype string) []manifest.Layer {
+	return slices.DeleteFunc(layers, func(layer manifest.Layer) bool {
 		if layer.MediaType != mediatype {
 			return false
 		}
@@ -680,7 +681,7 @@ func removeLayer(layers []Layer, mediatype string) []Layer {
 	})
 }
 
-func setTemplate(layers []Layer, t string) ([]Layer, error) {
+func setTemplate(layers []manifest.Layer, t string) ([]manifest.Layer, error) {
 	layers = removeLayer(layers, "application/vnd.ollama.image.template")
 	if _, err := template.Parse(t); err != nil {
 		return nil, fmt.Errorf("%w: %s", errBadTemplate, err)
@@ -690,7 +691,7 @@ func setTemplate(layers []Layer, t string) ([]Layer, error) {
 	}
 
 	blob := strings.NewReader(t)
-	layer, err := NewLayer(blob, "application/vnd.ollama.image.template")
+	layer, err := manifest.NewLayer(blob, "application/vnd.ollama.image.template")
 	if err != nil {
 		return nil, err
 	}
@@ -699,11 +700,11 @@ func setTemplate(layers []Layer, t string) ([]Layer, error) {
 	return layers, nil
 }
 
-func setSystem(layers []Layer, s string) ([]Layer, error) {
+func setSystem(layers []manifest.Layer, s string) ([]manifest.Layer, error) {
 	layers = removeLayer(layers, "application/vnd.ollama.image.system")
 	if s != "" {
 		blob := strings.NewReader(s)
-		layer, err := NewLayer(blob, "application/vnd.ollama.image.system")
+		layer, err := manifest.NewLayer(blob, "application/vnd.ollama.image.system")
 		if err != nil {
 			return nil, err
 		}
@@ -712,9 +713,9 @@ func setSystem(layers []Layer, s string) ([]Layer, error) {
 	return layers, nil
 }
 
-func setLicense(layers []Layer, l string) ([]Layer, error) {
+func setLicense(layers []manifest.Layer, l string) ([]manifest.Layer, error) {
 	blob := strings.NewReader(l)
-	layer, err := NewLayer(blob, "application/vnd.ollama.image.license")
+	layer, err := manifest.NewLayer(blob, "application/vnd.ollama.image.license")
 	if err != nil {
 		return nil, err
 	}
@@ -722,7 +723,7 @@ func setLicense(layers []Layer, l string) ([]Layer, error) {
 	return layers, nil
 }
 
-func setParameters(layers []Layer, p map[string]any) ([]Layer, error) {
+func setParameters(layers []manifest.Layer, p map[string]any) ([]manifest.Layer, error) {
 	if p == nil {
 		p = make(map[string]any)
 	}
@@ -731,7 +732,7 @@ func setParameters(layers []Layer, p map[string]any) ([]Layer, error) {
 			continue
 		}
 
-		digestPath, err := GetBlobsPath(layer.Digest)
+		digestPath, err := manifest.BlobsPath(layer.Digest)
 		if err != nil {
 			return nil, err
 		}
@@ -765,7 +766,7 @@ func setParameters(layers []Layer, p map[string]any) ([]Layer, error) {
 	if err := json.NewEncoder(&b).Encode(p); err != nil {
 		return nil, err
 	}
-	layer, err := NewLayer(&b, "application/vnd.ollama.image.params")
+	layer, err := manifest.NewLayer(&b, "application/vnd.ollama.image.params")
 	if err != nil {
 		return nil, err
 	}
@@ -773,7 +774,7 @@ func setParameters(layers []Layer, p map[string]any) ([]Layer, error) {
 	return layers, nil
 }
 
-func setMessages(layers []Layer, m []api.Message) ([]Layer, error) {
+func setMessages(layers []manifest.Layer, m []api.Message) ([]manifest.Layer, error) {
 	// this leaves the old messages intact if no new messages were specified
 	// which may not be the correct behaviour
 	if len(m) == 0 {
@@ -786,7 +787,7 @@ func setMessages(layers []Layer, m []api.Message) ([]Layer, error) {
 	if err := json.NewEncoder(&b).Encode(m); err != nil {
 		return nil, err
 	}
-	layer, err := NewLayer(&b, "application/vnd.ollama.image.messages")
+	layer, err := manifest.NewLayer(&b, "application/vnd.ollama.image.messages")
 	if err != nil {
 		return nil, err
 	}
@@ -794,7 +795,7 @@ func setMessages(layers []Layer, m []api.Message) ([]Layer, error) {
 	return layers, nil
 }
 
-func createConfigLayer(layers []Layer, config model.ConfigV2) (*Layer, error) {
+func createConfigLayer(layers []manifest.Layer, config model.ConfigV2) (*manifest.Layer, error) {
 	digests := make([]string, len(layers))
 	for i, layer := range layers {
 		digests[i] = layer.Digest
@@ -805,7 +806,7 @@ func createConfigLayer(layers []Layer, config model.ConfigV2) (*Layer, error) {
 	if err := json.NewEncoder(&b).Encode(config); err != nil {
 		return nil, err
 	}
-	layer, err := NewLayer(&b, "application/vnd.docker.container.image.v1+json")
+	layer, err := manifest.NewLayer(&b, "application/vnd.docker.container.image.v1+json")
 	if err != nil {
 		return nil, err
 	}

server/create_test.go

@@ -10,6 +10,7 @@ import (
 	"testing"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/manifest"
 )
 
 func TestConvertFromSafetensors(t *testing.T) {
@@ -17,7 +18,7 @@ func TestConvertFromSafetensors(t *testing.T) {
 
 	// Helper function to create a new layer and return its digest
 	makeTemp := func(content string) string {
-		l, err := NewLayer(strings.NewReader(content), "application/octet-stream")
+		l, err := manifest.NewLayer(strings.NewReader(content), "application/octet-stream")
 		if err != nil {
 			t.Fatalf("Failed to create layer: %v", err)
 		}

server/download.go

@@ -24,6 +24,8 @@ import (
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/manifest"
+	"github.com/ollama/ollama/types/model"
 )
 
 const maxRetries = 6
@@ -456,7 +458,7 @@ func (b *blobDownload) Wait(ctx context.Context, fn func(api.ProgressResponse))
 }
 
 type downloadOpts struct {
-	mp      ModelPath
+	n       model.Name
 	digest  string
 	regOpts *registryOptions
 	fn      func(api.ProgressResponse)
@@ -465,10 +467,10 @@ type downloadOpts struct {
 // downloadBlob downloads a blob from the registry and stores it in the blobs directory
 func downloadBlob(ctx context.Context, opts downloadOpts) (cacheHit bool, _ error) {
 	if opts.digest == "" {
-		return false, fmt.Errorf(("%s: %s"), opts.mp.GetNamespaceRepository(), "digest is empty")
+		return false, fmt.Errorf(("%s: %s"), opts.n.DisplayNamespaceModel(), "digest is empty")
 	}
 
-	fp, err := GetBlobsPath(opts.digest)
+	fp, err := manifest.BlobsPath(opts.digest)
 	if err != nil {
 		return false, err
 	}
@@ -492,8 +494,8 @@ func downloadBlob(ctx context.Context, opts downloadOpts) (cacheHit bool, _ erro
 	data, ok := blobDownloadManager.LoadOrStore(opts.digest, &blobDownload{Name: fp, Digest: opts.digest})
 	download := data.(*blobDownload)
 	if !ok {
-		requestURL := opts.mp.BaseURL()
-		requestURL = requestURL.JoinPath("v2", opts.mp.GetNamespaceRepository(), "blobs", opts.digest)
+		requestURL := opts.n.BaseURL()
+		requestURL = requestURL.JoinPath("v2", opts.n.DisplayNamespaceModel(), "blobs", opts.digest)
 		if err := download.Prepare(ctx, requestURL, opts.regOpts); err != nil {
 			blobDownloadManager.Delete(opts.digest)
 			return false, err

server/images.go

@@ -4,7 +4,6 @@ import (
 	"bytes"
 	"context"
 	"crypto/sha256"
-	"encoding/hex"
 	"encoding/json"
 	"errors"
 	"fmt"
@@ -24,6 +23,7 @@ import (
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/fs/gguf"
+	"github.com/ollama/ollama/manifest"
 	"github.com/ollama/ollama/model/parsers"
 	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/template"
@@ -274,44 +274,22 @@ func (m *Model) String() string {
 	return modelfile.String()
 }
 
-func GetManifest(mp ModelPath) (*Manifest, string, error) {
-	fp, err := mp.GetManifestPath()
-	if err != nil {
-		return nil, "", err
-	}
-
-	f, err := os.Open(fp)
-	if err != nil {
-		return nil, "", err
-	}
-	defer f.Close()
-
-	sha256sum := sha256.New()
-
-	var manifest Manifest
-	if err := json.NewDecoder(io.TeeReader(f, sha256sum)).Decode(&manifest); err != nil {
-		return nil, "", err
-	}
-
-	return &manifest, hex.EncodeToString(sha256sum.Sum(nil)), nil
-}
-
 func GetModel(name string) (*Model, error) {
-	mp := ParseModelPath(name)
-	manifest, digest, err := GetManifest(mp)
+	n := model.ParseName(name)
+	mf, err := manifest.ParseNamedManifest(n)
 	if err != nil {
 		return nil, err
 	}
 
-	model := &Model{
-		Name:      mp.GetFullTagname(),
-		ShortName: mp.GetShortTagname(),
-		Digest:    digest,
+	m := &Model{
+		Name:      n.String(),
+		ShortName: n.DisplayShortest(),
+		Digest:    mf.Digest(),
 		Template:  template.DefaultTemplate,
 	}
 
-	if manifest.Config.Digest != "" {
-		filename, err := GetBlobsPath(manifest.Config.Digest)
+	if mf.Config.Digest != "" {
+		filename, err := manifest.BlobsPath(mf.Config.Digest)
 		if err != nil {
 			return nil, err
 		}
@@ -322,29 +300,29 @@ func GetModel(name string) (*Model, error) {
 		}
 		defer configFile.Close()
 
-		if err := json.NewDecoder(configFile).Decode(&model.Config); err != nil {
+		if err := json.NewDecoder(configFile).Decode(&m.Config); err != nil {
 			return nil, err
 		}
 	}
 
-	for _, layer := range manifest.Layers {
-		filename, err := GetBlobsPath(layer.Digest)
+	for _, layer := range mf.Layers {
+		filename, err := manifest.BlobsPath(layer.Digest)
 		if err != nil {
 			return nil, err
 		}
 
 		switch layer.MediaType {
 		case "application/vnd.ollama.image.model":
-			model.ModelPath = filename
-			model.ParentModel = layer.From
+			m.ModelPath = filename
+			m.ParentModel = layer.From
 		case "application/vnd.ollama.image.embed":
 			// Deprecated in versions  > 0.1.2
 			// TODO: remove this warning in a future version
 			slog.Info("WARNING: model contains embeddings, but embeddings in modelfiles have been deprecated and will be ignored.")
 		case "application/vnd.ollama.image.adapter":
-			model.AdapterPaths = append(model.AdapterPaths, filename)
+			m.AdapterPaths = append(m.AdapterPaths, filename)
 		case "application/vnd.ollama.image.projector":
-			model.ProjectorPaths = append(model.ProjectorPaths, filename)
+			m.ProjectorPaths = append(m.ProjectorPaths, filename)
 		case "application/vnd.ollama.image.prompt",
 			"application/vnd.ollama.image.template":
 			bts, err := os.ReadFile(filename)
@@ -352,7 +330,7 @@ func GetModel(name string) (*Model, error) {
 				return nil, err
 			}
 
-			model.Template, err = template.Parse(string(bts))
+			m.Template, err = template.Parse(string(bts))
 			if err != nil {
 				return nil, err
 			}
@@ -362,7 +340,7 @@ func GetModel(name string) (*Model, error) {
 				return nil, err
 			}
 
-			model.System = string(bts)
+			m.System = string(bts)
 		case "application/vnd.ollama.image.params":
 			params, err := os.Open(filename)
 			if err != nil {
@@ -371,7 +349,7 @@ func GetModel(name string) (*Model, error) {
 			defer params.Close()
 
 			// parse model options parameters into a map so that we can see which fields have been specified explicitly
-			if err = json.NewDecoder(params).Decode(&model.Options); err != nil {
+			if err = json.NewDecoder(params).Decode(&m.Options); err != nil {
 				return nil, err
 			}
 		case "application/vnd.ollama.image.messages":
@@ -381,7 +359,7 @@ func GetModel(name string) (*Model, error) {
 			}
 			defer msgs.Close()
 
-			if err = json.NewDecoder(msgs).Decode(&model.Messages); err != nil {
+			if err = json.NewDecoder(msgs).Decode(&m.Messages); err != nil {
 				return nil, err
 			}
 		case "application/vnd.ollama.image.license":
@@ -389,11 +367,11 @@ func GetModel(name string) (*Model, error) {
 			if err != nil {
 				return nil, err
 			}
-			model.License = append(model.License, string(bts))
+			m.License = append(m.License, string(bts))
 		}
 	}
 
-	return model, nil
+	return m, nil
 }
 
 func CopyModel(src, dst model.Name) error {
@@ -408,7 +386,7 @@ func CopyModel(src, dst model.Name) error {
 		return nil
 	}
 
-	manifests, err := GetManifestPath()
+	manifests, err := manifest.Path()
 	if err != nil {
 		return err
 	}
@@ -437,7 +415,7 @@ func CopyModel(src, dst model.Name) error {
 
 func deleteUnusedLayers(deleteMap map[string]struct{}) error {
 	// Ignore corrupt manifests to avoid blocking deletion of layers that are freshly orphaned
-	manifests, err := Manifests(true)
+	manifests, err := manifest.Manifests(true)
 	if err != nil {
 		return err
 	}
@@ -452,7 +430,7 @@ func deleteUnusedLayers(deleteMap map[string]struct{}) error {
 
 	// only delete the files which are still in the deleteMap
 	for k := range deleteMap {
-		fp, err := GetBlobsPath(k)
+		fp, err := manifest.BlobsPath(k)
 		if err != nil {
 			slog.Info(fmt.Sprintf("couldn't get file path for '%s': %v", k, err))
 			continue
@@ -468,7 +446,7 @@ func deleteUnusedLayers(deleteMap map[string]struct{}) error {
 
 func PruneLayers() error {
 	deleteMap := make(map[string]struct{})
-	p, err := GetBlobsPath("")
+	p, err := manifest.BlobsPath("")
 	if err != nil {
 		return err
 	}
@@ -483,9 +461,9 @@ func PruneLayers() error {
 		name := blob.Name()
 		name = strings.ReplaceAll(name, "-", ":")
 
-		_, err := GetBlobsPath(name)
+		_, err := manifest.BlobsPath(name)
 		if err != nil {
-			if errors.Is(err, ErrInvalidDigestFormat) {
+			if errors.Is(err, manifest.ErrInvalidDigestFormat) {
 				// remove invalid blobs (e.g. partial downloads)
 				if err := os.Remove(filepath.Join(p, blob.Name())); err != nil {
 					slog.Error("couldn't remove blob", "blob", blob.Name(), "error", err)
@@ -510,63 +488,30 @@ func PruneLayers() error {
 	return nil
 }
 
-func PruneDirectory(path string) error {
-	info, err := os.Lstat(path)
-	if err != nil {
-		return err
-	}
-
-	if info.IsDir() && info.Mode()&os.ModeSymlink == 0 {
-		entries, err := os.ReadDir(path)
-		if err != nil {
-			return err
-		}
-
-		for _, entry := range entries {
-			if err := PruneDirectory(filepath.Join(path, entry.Name())); err != nil {
-				return err
-			}
-		}
-
-		entries, err = os.ReadDir(path)
-		if err != nil {
-			return err
-		}
-
-		if len(entries) > 0 {
-			return nil
-		}
-
-		return os.Remove(path)
-	}
-
-	return nil
-}
-
 func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
-	mp := ParseModelPath(name)
+	n := model.ParseName(name)
 	fn(api.ProgressResponse{Status: "retrieving manifest"})
 
-	if mp.ProtocolScheme == "http" && !regOpts.Insecure {
+	if n.ProtocolScheme == "http" && !regOpts.Insecure {
 		return errInsecureProtocol
 	}
 
-	manifest, _, err := GetManifest(mp)
+	mf, err := manifest.ParseNamedManifest(n)
 	if err != nil {
 		fn(api.ProgressResponse{Status: "couldn't retrieve manifest"})
 		return err
 	}
 
-	var layers []Layer
-	layers = append(layers, manifest.Layers...)
-	if manifest.Config.Digest != "" {
-		layers = append(layers, manifest.Config)
+	var layers []manifest.Layer
+	layers = append(layers, mf.Layers...)
+	if mf.Config.Digest != "" {
+		layers = append(layers, mf.Config)
 	}
 
 	// Use fast transfer for models with tensor layers (many small blobs)
 	if hasTensorLayers(layers) {
 		// Read raw manifest JSON to preserve tensor metadata fields
-		manifestPath, err := mp.GetManifestPath()
+		manifestPath, err := manifest.PathForName(n)
 		if err != nil {
 			return err
 		}
@@ -574,7 +519,7 @@ func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 		if err != nil {
 			return err
 		}
-		if err := pushWithTransfer(ctx, mp, layers, manifestJSON, regOpts, fn); err != nil {
+		if err := pushWithTransfer(ctx, n, layers, manifestJSON, regOpts, fn); err != nil {
 			return err
 		}
 		fn(api.ProgressResponse{Status: "success"})
@@ -582,17 +527,17 @@ func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	}
 
 	for _, layer := range layers {
-		if err := uploadBlob(ctx, mp, layer, regOpts, fn); err != nil {
+		if err := uploadBlob(ctx, n, layer, regOpts, fn); err != nil {
 			slog.Info(fmt.Sprintf("error uploading blob: %v", err))
 			return err
 		}
 	}
 
 	fn(api.ProgressResponse{Status: "pushing manifest"})
-	requestURL := mp.BaseURL()
-	requestURL = requestURL.JoinPath("v2", mp.GetNamespaceRepository(), "manifests", mp.Tag)
+	requestURL := n.BaseURL()
+	requestURL = requestURL.JoinPath("v2", n.DisplayNamespaceModel(), "manifests", n.Tag)
 
-	manifestJSON, err := json.Marshal(manifest)
+	manifestJSON, err := json.Marshal(mf)
 	if err != nil {
 		return err
 	}
@@ -611,44 +556,44 @@ func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 }
 
 func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
-	mp := ParseModelPath(name)
+	n := model.ParseName(name)
 
 	// build deleteMap to prune unused layers
 	deleteMap := make(map[string]struct{})
-	manifest, _, err := GetManifest(mp)
+	existingMf, err := manifest.ParseNamedManifest(n)
 	if errors.Is(err, os.ErrNotExist) {
 		// noop
 	} else if err != nil {
 		slog.Warn("pulling model with bad existing manifest", "name", name, "error", err)
 	} else {
-		for _, l := range manifest.Layers {
+		for _, l := range existingMf.Layers {
 			deleteMap[l.Digest] = struct{}{}
 		}
-		if manifest.Config.Digest != "" {
-			deleteMap[manifest.Config.Digest] = struct{}{}
+		if existingMf.Config.Digest != "" {
+			deleteMap[existingMf.Config.Digest] = struct{}{}
 		}
 	}
 
-	if mp.ProtocolScheme == "http" && !regOpts.Insecure {
+	if n.ProtocolScheme == "http" && !regOpts.Insecure {
 		return errInsecureProtocol
 	}
 
 	fn(api.ProgressResponse{Status: "pulling manifest"})
 
-	manifest, err = pullModelManifest(ctx, mp, regOpts)
+	mf, err := pullModelManifest(ctx, n, regOpts)
 	if err != nil {
 		return fmt.Errorf("pull model manifest: %s", err)
 	}
 
-	var layers []Layer
-	layers = append(layers, manifest.Layers...)
-	if manifest.Config.Digest != "" {
-		layers = append(layers, manifest.Config)
+	var layers []manifest.Layer
+	layers = append(layers, mf.Layers...)
+	if mf.Config.Digest != "" {
+		layers = append(layers, mf.Config)
 	}
 
 	// Use fast transfer for models with tensor layers (many small blobs)
 	if hasTensorLayers(layers) {
-		if err := pullWithTransfer(ctx, mp, layers, manifest, regOpts, fn); err != nil {
+		if err := pullWithTransfer(ctx, n, layers, mf, regOpts, fn); err != nil {
 			return err
 		}
 		fn(api.ProgressResponse{Status: "success"})
@@ -658,7 +603,7 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	skipVerify := make(map[string]bool)
 	for _, layer := range layers {
 		cacheHit, err := downloadBlob(ctx, downloadOpts{
-			mp:      mp,
+			n:       n,
 			digest:  layer.Digest,
 			regOpts: regOpts,
 			fn:      fn,
@@ -677,7 +622,7 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 		}
 		if err := verifyBlob(layer.Digest); err != nil {
 			if errors.Is(err, errDigestMismatch) {
-				fp, err := GetBlobsPath(layer.Digest)
+				fp, err := manifest.BlobsPath(layer.Digest)
 				if err != nil {
 					return err
 				}
@@ -692,16 +637,16 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	for _, layer := range layers {
 		delete(deleteMap, layer.Digest)
 	}
-	delete(deleteMap, manifest.Config.Digest)
+	delete(deleteMap, mf.Config.Digest)
 
 	fn(api.ProgressResponse{Status: "writing manifest"})
 
-	manifestJSON, err := json.Marshal(manifest)
+	manifestJSON, err := json.Marshal(mf)
 	if err != nil {
 		return err
 	}
 
-	fp, err := mp.GetManifestPath()
+	fp, err := manifest.PathForName(n)
 	if err != nil {
 		return err
 	}
@@ -728,9 +673,9 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 }
 
 // hasTensorLayers checks if any layer has tensor media type.
-func hasTensorLayers(layers []Layer) bool {
+func hasTensorLayers(layers []manifest.Layer) bool {
 	for _, layer := range layers {
-		if layer.MediaType == MediaTypeImageTensor {
+		if layer.MediaType == manifest.MediaTypeImageTensor {
 			return true
 		}
 	}
@@ -738,7 +683,7 @@ func hasTensorLayers(layers []Layer) bool {
 }
 
 // pullWithTransfer uses the simplified x/transfer package for downloading blobs.
-func pullWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifest *Manifest, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
+func pullWithTransfer(ctx context.Context, n model.Name, layers []manifest.Layer, mf *manifest.Manifest, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
 	blobs := make([]transfer.Blob, len(layers))
 	for i, layer := range layers {
 		blobs[i] = transfer.Blob{
@@ -747,12 +692,12 @@ func pullWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifes
 		}
 	}
 
-	destDir, err := GetBlobsPath("")
+	destDir, err := manifest.BlobsPath("")
 	if err != nil {
 		return err
 	}
 
-	base := mp.BaseURL()
+	base := n.BaseURL()
 	if base.Scheme != "http" && regOpts != nil && regOpts.Insecure {
 		base.Scheme = "http"
 	}
@@ -784,7 +729,7 @@ func pullWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifes
 		Blobs:      blobs,
 		BaseURL:    baseURL,
 		DestDir:    destDir,
-		Repository: mp.GetNamespaceRepository(),
+		Repository: n.DisplayNamespaceModel(),
 		Progress:   progress,
 		Token:      regOpts.Token,
 		GetToken:   getToken,
@@ -795,12 +740,12 @@ func pullWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifes
 
 	// Write manifest
 	fn(api.ProgressResponse{Status: "writing manifest"})
-	manifestJSON, err := json.Marshal(manifest)
+	manifestJSON, err := json.Marshal(mf)
 	if err != nil {
 		return err
 	}
 
-	fp, err := mp.GetManifestPath()
+	fp, err := manifest.PathForName(n)
 	if err != nil {
 		return err
 	}
@@ -812,7 +757,7 @@ func pullWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifes
 }
 
 // pushWithTransfer uses the simplified x/transfer package for uploading blobs and manifest.
-func pushWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifestJSON []byte, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
+func pushWithTransfer(ctx context.Context, n model.Name, layers []manifest.Layer, manifestJSON []byte, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
 	blobs := make([]transfer.Blob, len(layers))
 	for i, layer := range layers {
 		blobs[i] = transfer.Blob{
@@ -822,12 +767,12 @@ func pushWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifes
 		}
 	}
 
-	srcDir, err := GetBlobsPath("")
+	srcDir, err := manifest.BlobsPath("")
 	if err != nil {
 		return err
 	}
 
-	base := mp.BaseURL()
+	base := n.BaseURL()
 	if base.Scheme != "http" && regOpts != nil && regOpts.Insecure {
 		base.Scheme = "http"
 	}
@@ -864,13 +809,13 @@ func pushWithTransfer(ctx context.Context, mp ModelPath, layers []Layer, manifes
 		GetToken:    getToken,
 		Logger:      slog.Default(),
 		Manifest:    manifestJSON,
-		ManifestRef: mp.Tag,
-		Repository:  mp.GetNamespaceRepository(),
+		ManifestRef: n.Tag,
+		Repository:  n.DisplayNamespaceModel(),
 	})
 }
 
-func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptions) (*Manifest, error) {
-	requestURL := mp.BaseURL().JoinPath("v2", mp.GetNamespaceRepository(), "manifests", mp.Tag)
+func pullModelManifest(ctx context.Context, n model.Name, regOpts *registryOptions) (*manifest.Manifest, error) {
+	requestURL := n.BaseURL().JoinPath("v2", n.DisplayNamespaceModel(), "manifests", n.Tag)
 
 	headers := make(http.Header)
 	headers.Set("Accept", "application/vnd.docker.distribution.manifest.v2+json")
@@ -880,7 +825,7 @@ func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptio
 	}
 	defer resp.Body.Close()
 
-	var m Manifest
+	var m manifest.Manifest
 	if err := json.NewDecoder(resp.Body).Decode(&m); err != nil {
 		return nil, err
 	}
@@ -1042,7 +987,7 @@ func parseRegistryChallenge(authStr string) registryChallenge {
 var errDigestMismatch = errors.New("digest mismatch, file must be downloaded again")
 
 func verifyBlob(digest string) error {
-	fp, err := GetBlobsPath(digest)
+	fp, err := manifest.BlobsPath(digest)
 	if err != nil {
 		return err
 	}

server/model.go

@@ -13,6 +13,7 @@ import (
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/manifest"
 	"github.com/ollama/ollama/template"
 	"github.com/ollama/ollama/types/model"
 )
@@ -20,19 +21,19 @@ import (
 var intermediateBlobs map[string]string = make(map[string]string)
 
 type layerGGML struct {
-	Layer
+	manifest.Layer
 	*ggml.GGML
 }
 
 func parseFromModel(ctx context.Context, name model.Name, fn func(api.ProgressResponse)) (layers []*layerGGML, err error) {
-	m, err := ParseNamedManifest(name)
+	m, err := manifest.ParseNamedManifest(name)
 	switch {
 	case errors.Is(err, os.ErrNotExist):
 		if err := PullModel(ctx, name.String(), &registryOptions{}, fn); err != nil {
 			return nil, err
 		}
 
-		m, err = ParseNamedManifest(name)
+		m, err = manifest.ParseNamedManifest(name)
 		if err != nil {
 			return nil, err
 		}
@@ -41,7 +42,7 @@ func parseFromModel(ctx context.Context, name model.Name, fn func(api.ProgressRe
 	}
 
 	for _, layer := range m.Layers {
-		layer, err := NewLayerFromLayer(layer.Digest, layer.MediaType, name.DisplayShortest())
+		layer, err := manifest.NewLayerFromLayer(layer.Digest, layer.MediaType, name.DisplayShortest())
 		if err != nil {
 			return nil, err
 		}
@@ -50,7 +51,7 @@ func parseFromModel(ctx context.Context, name model.Name, fn func(api.ProgressRe
 		case "application/vnd.ollama.image.model",
 			"application/vnd.ollama.image.projector",
 			"application/vnd.ollama.image.adapter":
-			blobpath, err := GetBlobsPath(layer.Digest)
+			blobpath, err := manifest.BlobsPath(layer.Digest)
 			if err != nil {
 				return nil, err
 			}
@@ -81,12 +82,12 @@ func detectChatTemplate(layers []*layerGGML) ([]*layerGGML, error) {
 			if t, err := template.Named(s); err != nil {
 				slog.Debug("template detection", "error", err, "template", s)
 			} else {
-				layer, err := NewLayer(t.Reader(), "application/vnd.ollama.image.template")
+				layer, err := manifest.NewLayer(t.Reader(), "application/vnd.ollama.image.template")
 				if err != nil {
 					return nil, err
 				}
 
-				layer.status = fmt.Sprintf("using autodetected template %s", t.Name)
+				layer.Status = fmt.Sprintf("using autodetected template %s", t.Name)
 				layers = append(layers, &layerGGML{layer, nil})
 
 				if t.Parameters != nil {
@@ -95,7 +96,7 @@ func detectChatTemplate(layers []*layerGGML) ([]*layerGGML, error) {
 						return nil, err
 					}
 
-					layer, err := NewLayer(&b, "application/vnd.ollama.image.params")
+					layer, err := manifest.NewLayer(&b, "application/vnd.ollama.image.params")
 					if err != nil {
 						return nil, err
 					}

server/modelpath.go

@@ -1,146 +0,0 @@
-package server
-
-import (
-	"errors"
-	"fmt"
-	"io/fs"
-	"net/url"
-	"os"
-	"path/filepath"
-	"regexp"
-	"strings"
-
-	"github.com/ollama/ollama/envconfig"
-	"github.com/ollama/ollama/types/model"
-)
-
-type ModelPath struct {
-	ProtocolScheme string
-	Registry       string
-	Namespace      string
-	Repository     string
-	Tag            string
-}
-
-const (
-	DefaultRegistry       = "registry.ollama.ai"
-	DefaultNamespace      = "library"
-	DefaultTag            = "latest"
-	DefaultProtocolScheme = "https"
-)
-
-var (
-	ErrInvalidImageFormat  = errors.New("invalid image format")
-	ErrInvalidDigestFormat = errors.New("invalid digest format")
-	ErrInvalidProtocol     = errors.New("invalid protocol scheme")
-	ErrInsecureProtocol    = errors.New("insecure protocol http")
-	ErrModelPathInvalid    = errors.New("invalid model path")
-)
-
-func ParseModelPath(name string) ModelPath {
-	mp := ModelPath{
-		ProtocolScheme: DefaultProtocolScheme,
-		Registry:       DefaultRegistry,
-		Namespace:      DefaultNamespace,
-		Repository:     "",
-		Tag:            DefaultTag,
-	}
-
-	before, after, found := strings.Cut(name, "://")
-	if found {
-		mp.ProtocolScheme = before
-		name = after
-	}
-
-	name = strings.ReplaceAll(name, string(os.PathSeparator), "/")
-	parts := strings.Split(name, "/")
-	switch len(parts) {
-	case 3:
-		mp.Registry = parts[0]
-		mp.Namespace = parts[1]
-		mp.Repository = parts[2]
-	case 2:
-		mp.Namespace = parts[0]
-		mp.Repository = parts[1]
-	case 1:
-		mp.Repository = parts[0]
-	}
-
-	if repo, tag, found := strings.Cut(mp.Repository, ":"); found {
-		mp.Repository = repo
-		mp.Tag = tag
-	}
-
-	return mp
-}
-
-func (mp ModelPath) GetNamespaceRepository() string {
-	return fmt.Sprintf("%s/%s", mp.Namespace, mp.Repository)
-}
-
-func (mp ModelPath) GetFullTagname() string {
-	return fmt.Sprintf("%s/%s/%s:%s", mp.Registry, mp.Namespace, mp.Repository, mp.Tag)
-}
-
-func (mp ModelPath) GetShortTagname() string {
-	if mp.Registry == DefaultRegistry {
-		if mp.Namespace == DefaultNamespace {
-			return fmt.Sprintf("%s:%s", mp.Repository, mp.Tag)
-		}
-		return fmt.Sprintf("%s/%s:%s", mp.Namespace, mp.Repository, mp.Tag)
-	}
-	return fmt.Sprintf("%s/%s/%s:%s", mp.Registry, mp.Namespace, mp.Repository, mp.Tag)
-}
-
-// GetManifestPath returns the path to the manifest file for the given model path, it is up to the caller to create the directory if it does not exist.
-func (mp ModelPath) GetManifestPath() (string, error) {
-	name := model.Name{
-		Host:      mp.Registry,
-		Namespace: mp.Namespace,
-		Model:     mp.Repository,
-		Tag:       mp.Tag,
-	}
-	if !name.IsValid() {
-		return "", fs.ErrNotExist
-	}
-	return filepath.Join(envconfig.Models(), "manifests", name.Filepath()), nil
-}
-
-func (mp ModelPath) BaseURL() *url.URL {
-	return &url.URL{
-		Scheme: mp.ProtocolScheme,
-		Host:   mp.Registry,
-	}
-}
-
-func GetManifestPath() (string, error) {
-	path := filepath.Join(envconfig.Models(), "manifests")
-	if err := os.MkdirAll(path, 0o755); err != nil {
-		return "", fmt.Errorf("%w: ensure path elements are traversable", err)
-	}
-
-	return path, nil
-}
-
-func GetBlobsPath(digest string) (string, error) {
-	// only accept actual sha256 digests
-	pattern := "^sha256[:-][0-9a-fA-F]{64}$"
-	re := regexp.MustCompile(pattern)
-
-	if digest != "" && !re.MatchString(digest) {
-		return "", ErrInvalidDigestFormat
-	}
-
-	digest = strings.ReplaceAll(digest, ":", "-")
-	path := filepath.Join(envconfig.Models(), "blobs", digest)
-	dirPath := filepath.Dir(path)
-	if digest == "" {
-		dirPath = path
-	}
-
-	if err := os.MkdirAll(dirPath, 0o755); err != nil {
-		return "", fmt.Errorf("%w: ensure path elements are traversable", err)
-	}
-
-	return path, nil
-}

server/modelpath_test.go

@@ -1,153 +0,0 @@
-package server
-
-import (
-	"path/filepath"
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-func TestGetBlobsPath(t *testing.T) {
-	// GetBlobsPath expects an actual directory to exist
-	tempDir := t.TempDir()
-
-	tests := []struct {
-		name     string
-		digest   string
-		expected string
-		err      error
-	}{
-		{
-			"empty digest",
-			"",
-			filepath.Join(tempDir, "blobs"),
-			nil,
-		},
-		{
-			"valid with colon",
-			"sha256:456402914e838a953e0cf80caa6adbe75383d9e63584a964f504a7bbb8f7aad9",
-			filepath.Join(tempDir, "blobs", "sha256-456402914e838a953e0cf80caa6adbe75383d9e63584a964f504a7bbb8f7aad9"),
-			nil,
-		},
-		{
-			"valid with dash",
-			"sha256-456402914e838a953e0cf80caa6adbe75383d9e63584a964f504a7bbb8f7aad9",
-			filepath.Join(tempDir, "blobs", "sha256-456402914e838a953e0cf80caa6adbe75383d9e63584a964f504a7bbb8f7aad9"),
-			nil,
-		},
-		{
-			"digest too short",
-			"sha256-45640291",
-			"",
-			ErrInvalidDigestFormat,
-		},
-		{
-			"digest too long",
-			"sha256-456402914e838a953e0cf80caa6adbe75383d9e63584a964f504a7bbb8f7aad9aaaaaaaaaa",
-			"",
-			ErrInvalidDigestFormat,
-		},
-		{
-			"digest invalid chars",
-			"../sha256-456402914e838a953e0cf80caa6adbe75383d9e63584a964f504a7bbb8f7a",
-			"",
-			ErrInvalidDigestFormat,
-		},
-	}
-	for _, tc := range tests {
-		t.Run(tc.name, func(t *testing.T) {
-			t.Setenv("OLLAMA_MODELS", tempDir)
-
-			got, err := GetBlobsPath(tc.digest)
-
-			require.ErrorIs(t, tc.err, err, tc.name)
-			assert.Equal(t, tc.expected, got, tc.name)
-		})
-	}
-}
-
-func TestParseModelPath(t *testing.T) {
-	tests := []struct {
-		name string
-		arg  string
-		want ModelPath
-	}{
-		{
-			"full path https",
-			"https://example.com/ns/repo:tag",
-			ModelPath{
-				ProtocolScheme: "https",
-				Registry:       "example.com",
-				Namespace:      "ns",
-				Repository:     "repo",
-				Tag:            "tag",
-			},
-		},
-		{
-			"full path http",
-			"http://example.com/ns/repo:tag",
-			ModelPath{
-				ProtocolScheme: "http",
-				Registry:       "example.com",
-				Namespace:      "ns",
-				Repository:     "repo",
-				Tag:            "tag",
-			},
-		},
-		{
-			"no protocol",
-			"example.com/ns/repo:tag",
-			ModelPath{
-				ProtocolScheme: "https",
-				Registry:       "example.com",
-				Namespace:      "ns",
-				Repository:     "repo",
-				Tag:            "tag",
-			},
-		},
-		{
-			"no registry",
-			"ns/repo:tag",
-			ModelPath{
-				ProtocolScheme: "https",
-				Registry:       DefaultRegistry,
-				Namespace:      "ns",
-				Repository:     "repo",
-				Tag:            "tag",
-			},
-		},
-		{
-			"no namespace",
-			"repo:tag",
-			ModelPath{
-				ProtocolScheme: "https",
-				Registry:       DefaultRegistry,
-				Namespace:      DefaultNamespace,
-				Repository:     "repo",
-				Tag:            "tag",
-			},
-		},
-		{
-			"no tag",
-			"repo",
-			ModelPath{
-				ProtocolScheme: "https",
-				Registry:       DefaultRegistry,
-				Namespace:      DefaultNamespace,
-				Repository:     "repo",
-				Tag:            DefaultTag,
-			},
-		},
-	}
-
-	for _, tc := range tests {
-		t.Run(tc.name, func(t *testing.T) {
-			got := ParseModelPath(tc.arg)
-
-			if got != tc.want {
-				t.Errorf("got: %q want: %q", got, tc.want)
-			}
-		})
-	}
-}

server/routes.go

@@ -39,6 +39,7 @@ import (
 	"github.com/ollama/ollama/fs/ggml"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/logutil"
+	"github.com/ollama/ollama/manifest"
 	"github.com/ollama/ollama/middleware"
 	"github.com/ollama/ollama/model/parsers"
 	"github.com/ollama/ollama/model/renderers"
@@ -974,7 +975,7 @@ func (s *Server) PushHandler(c *gin.Context) {
 // is.
 func getExistingName(n model.Name) (model.Name, error) {
 	var zero model.Name
-	existing, err := Manifests(true)
+	existing, err := manifest.Manifests(true)
 	if err != nil {
 		return zero, err
 	}
@@ -1018,7 +1019,7 @@ func (s *Server) DeleteHandler(c *gin.Context) {
 		return
 	}
 
-	m, err := ParseNamedManifest(n)
+	m, err := manifest.ParseNamedManifest(n)
 	if err != nil {
 		switch {
 		case os.IsNotExist(err):
@@ -1080,7 +1081,7 @@ func (s *Server) ShowHandler(c *gin.Context) {
 func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 	name := model.ParseName(req.Model)
 	if !name.IsValid() {
-		return nil, ErrModelPathInvalid
+		return nil, model.Unqualified(name)
 	}
 	name, err := getExistingName(name)
 	if err != nil {
@@ -1112,7 +1113,7 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 
 	// For safetensors LLM models (experimental), populate details from config.json
 	if m.Config.ModelFormat == "safetensors" && slices.Contains(m.Config.Capabilities, "completion") {
-		if info, err := xserver.GetSafetensorsLLMInfo(name.String()); err == nil {
+		if info, err := xserver.GetSafetensorsLLMInfo(name); err == nil {
 			if arch, ok := info["general.architecture"].(string); ok && arch != "" {
 				modelDetails.Family = arch
 			}
@@ -1121,7 +1122,7 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 			}
 		}
 		// Get torch_dtype directly from config.json for quantization level
-		if dtype, err := xserver.GetSafetensorsDtype(name.String()); err == nil && dtype != "" {
+		if dtype, err := xserver.GetSafetensorsDtype(name); err == nil && dtype != "" {
 			modelDetails.QuantizationLevel = dtype
 		}
 	}
@@ -1135,7 +1136,7 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 		msgs[i] = api.Message{Role: msg.Role, Content: msg.Content}
 	}
 
-	manifest, err := ParseNamedManifest(name)
+	mf, err := manifest.ParseNamedManifest(name)
 	if err != nil {
 		return nil, err
 	}
@@ -1147,7 +1148,7 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 		Details:      modelDetails,
 		Messages:     msgs,
 		Capabilities: m.Capabilities(),
-		ModifiedAt:   manifest.fi.ModTime(),
+		ModifiedAt:   mf.FileInfo().ModTime(),
 		Requires:     m.Config.Requires,
 		// Several integrations crash on a nil/omitempty+empty ModelInfo, so by
 		// default we return an empty map.
@@ -1214,7 +1215,7 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 	if slices.Contains(m.Capabilities(), model.CapabilityImage) {
 		// Populate tensor info if verbose
 		if req.Verbose {
-			if tensors, err := xserver.GetSafetensorsTensorInfo(name.String()); err == nil {
+			if tensors, err := xserver.GetSafetensorsTensorInfo(name); err == nil {
 				resp.Tensors = tensors
 			}
 		}
@@ -1223,12 +1224,12 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 
 	// For safetensors LLM models (experimental), populate ModelInfo from config.json
 	if m.Config.ModelFormat == "safetensors" && slices.Contains(m.Config.Capabilities, "completion") {
-		if info, err := xserver.GetSafetensorsLLMInfo(name.String()); err == nil {
+		if info, err := xserver.GetSafetensorsLLMInfo(name); err == nil {
 			resp.ModelInfo = info
 		}
 		// Populate tensor info if verbose
 		if req.Verbose {
-			if tensors, err := xserver.GetSafetensorsTensorInfo(name.String()); err == nil {
+			if tensors, err := xserver.GetSafetensorsTensorInfo(name); err == nil {
 				resp.Tensors = tensors
 			}
 		}
@@ -1285,7 +1286,7 @@ func getModelData(digest string, verbose bool) (ggml.KV, ggml.Tensors, error) {
 }
 
 func (s *Server) ListHandler(c *gin.Context) {
-	ms, err := Manifests(true)
+	ms, err := manifest.Manifests(true)
 	if err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
@@ -1316,8 +1317,8 @@ func (s *Server) ListHandler(c *gin.Context) {
 			RemoteModel: cf.RemoteModel,
 			RemoteHost:  cf.RemoteHost,
 			Size:        m.Size(),
-			Digest:      m.digest,
-			ModifiedAt:  m.fi.ModTime(),
+			Digest:      m.Digest(),
+			ModifiedAt:  m.FileInfo().ModTime(),
 			Details: api.ModelDetails{
 				Format:            cf.ModelFormat,
 				Family:            cf.ModelFamily,
@@ -1376,7 +1377,7 @@ func (s *Server) CopyHandler(c *gin.Context) {
 }
 
 func (s *Server) HeadBlobHandler(c *gin.Context) {
-	path, err := GetBlobsPath(c.Param("digest"))
+	path, err := manifest.BlobsPath(c.Param("digest"))
 	if err != nil {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
@@ -1392,7 +1393,7 @@ func (s *Server) HeadBlobHandler(c *gin.Context) {
 
 func (s *Server) CreateBlobHandler(c *gin.Context) {
 	if ib, ok := intermediateBlobs[c.Param("digest")]; ok {
-		p, err := GetBlobsPath(ib)
+		p, err := manifest.BlobsPath(ib)
 		if err != nil {
 			c.AbortWithStatusJSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 			return
@@ -1410,7 +1411,7 @@ func (s *Server) CreateBlobHandler(c *gin.Context) {
 		}
 	}
 
-	path, err := GetBlobsPath(c.Param("digest"))
+	path, err := manifest.BlobsPath(c.Param("digest"))
 	if err != nil {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
@@ -1428,7 +1429,7 @@ func (s *Server) CreateBlobHandler(c *gin.Context) {
 		return
 	}
 
-	layer, err := NewLayer(c.Request.Body, "")
+	layer, err := manifest.NewLayer(c.Request.Body, "")
 	if err != nil {
 		c.AbortWithStatusJSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
@@ -1628,7 +1629,7 @@ func Serve(ln net.Listener) error {
 	slog.SetDefault(logutil.NewLogger(os.Stderr, envconfig.LogLevel()))
 	slog.Info("server config", "env", envconfig.Values())
 
-	blobsDir, err := GetBlobsPath("")
+	blobsDir, err := manifest.BlobsPath("")
 	if err != nil {
 		return err
 	}
@@ -1637,7 +1638,7 @@ func Serve(ln net.Listener) error {
 	}
 
 	if !envconfig.NoPrune() {
-		if _, err := Manifests(false); err != nil {
+		if _, err := manifest.Manifests(false); err != nil {
 			slog.Warn("corrupt manifests detected, skipping prune operation.  Re-pull or delete to clear", "error", err)
 		} else {
 			// clean up unused layers and manifests
@@ -1645,12 +1646,12 @@ func Serve(ln net.Listener) error {
 				return err
 			}
 
-			manifestsPath, err := GetManifestPath()
+			manifestsPath, err := manifest.Path()
 			if err != nil {
 				return err
 			}
 
-			if err := PruneDirectory(manifestsPath); err != nil {
+			if err := manifest.PruneDirectory(manifestsPath); err != nil {
 				return err
 			}
 		}

server/routes_create_test.go

@@ -25,6 +25,7 @@ import (
 	"github.com/ollama/ollama/convert"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/manifest"
 	"github.com/ollama/ollama/types/model"
 )
 
@@ -223,15 +224,15 @@ func TestCreateFromModelInheritsRendererParser(t *testing.T) {
 		t.Fatalf("expected status code 200, actual %d", w.Code)
 	}
 
-	manifest, err := ParseNamedManifest(model.ParseName("child"))
+	mf, err := manifest.ParseNamedManifest(model.ParseName("child"))
 	if err != nil {
 		t.Fatalf("parse manifest: %v", err)
 	}
-	if manifest.Config.Digest == "" {
+	if mf.Config.Digest == "" {
 		t.Fatalf("unexpected empty config digest for child manifest")
 	}
 
-	configPath, err := GetBlobsPath(manifest.Config.Digest)
+	configPath, err := manifest.BlobsPath(mf.Config.Digest)
 	if err != nil {
 		t.Fatalf("config blob path: %v", err)
 	}

server/routes_delete_test.go

@@ -10,6 +10,7 @@ import (
 	"github.com/gin-gonic/gin"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/manifest"
 	"github.com/ollama/ollama/types/model"
 )
 
@@ -93,13 +94,13 @@ func TestDeleteDuplicateLayers(t *testing.T) {
 		t.Fatal(err)
 	}
 
-	config, err := NewLayer(&b, "application/vnd.docker.container.image.v1+json")
+	config, err := manifest.NewLayer(&b, "application/vnd.docker.container.image.v1+json")
 	if err != nil {
 		t.Fatal(err)
 	}
 
 	// create a manifest with duplicate layers
-	if err := WriteManifest(n, config, []Layer{config}); err != nil {
+	if err := manifest.WriteManifest(n, config, []manifest.Layer{config}); err != nil {
 		t.Fatal(err)
 	}
 

server/upload.go

@@ -21,12 +21,14 @@ import (
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/manifest"
+	"github.com/ollama/ollama/types/model"
 )
 
 var blobUploadManager sync.Map
 
 type blobUpload struct {
-	Layer
+	manifest.Layer
 
 	Total     int64
 	Completed atomic.Int64
@@ -51,7 +53,7 @@ const (
 )
 
 func (b *blobUpload) Prepare(ctx context.Context, requestURL *url.URL, opts *registryOptions) error {
-	p, err := GetBlobsPath(b.Digest)
+	p, err := manifest.BlobsPath(b.Digest)
 	if err != nil {
 		return err
 	}
@@ -59,7 +61,7 @@ func (b *blobUpload) Prepare(ctx context.Context, requestURL *url.URL, opts *reg
 	if b.From != "" {
 		values := requestURL.Query()
 		values.Add("mount", b.Digest)
-		values.Add("from", ParseModelPath(b.From).GetNamespaceRepository())
+		values.Add("from", model.ParseName(b.From).DisplayNamespaceModel())
 		requestURL.RawQuery = values.Encode()
 	}
 
@@ -128,7 +130,7 @@ func (b *blobUpload) Run(ctx context.Context, opts *registryOptions) {
 	defer blobUploadManager.Delete(b.Digest)
 	ctx, b.CancelFunc = context.WithCancel(ctx)
 
-	p, err := GetBlobsPath(b.Digest)
+	p, err := manifest.BlobsPath(b.Digest)
 	if err != nil {
 		b.err = err
 		return
@@ -364,9 +366,9 @@ func (p *progressWriter) Rollback() {
 	p.written = 0
 }
 
-func uploadBlob(ctx context.Context, mp ModelPath, layer Layer, opts *registryOptions, fn func(api.ProgressResponse)) error {
-	requestURL := mp.BaseURL()
-	requestURL = requestURL.JoinPath("v2", mp.GetNamespaceRepository(), "blobs", layer.Digest)
+func uploadBlob(ctx context.Context, n model.Name, layer manifest.Layer, opts *registryOptions, fn func(api.ProgressResponse)) error {
+	requestURL := n.BaseURL()
+	requestURL = requestURL.JoinPath("v2", n.DisplayNamespaceModel(), "blobs", layer.Digest)
 
 	resp, err := makeRequestWithRetry(ctx, http.MethodHead, requestURL, nil, nil, opts)
 	switch {
@@ -388,8 +390,8 @@ func uploadBlob(ctx context.Context, mp ModelPath, layer Layer, opts *registryOp
 	data, ok := blobUploadManager.LoadOrStore(layer.Digest, &blobUpload{Layer: layer})
 	upload := data.(*blobUpload)
 	if !ok {
-		requestURL := mp.BaseURL()
-		requestURL = requestURL.JoinPath("v2", mp.GetNamespaceRepository(), "blobs/uploads/")
+		requestURL := n.BaseURL()
+		requestURL = requestURL.JoinPath("v2", n.DisplayNamespaceModel(), "blobs/uploads/")
 		if err := upload.Prepare(ctx, requestURL, opts); err != nil {
 			blobUploadManager.Delete(layer.Digest)
 			return err

types/model/name.go

@@ -7,6 +7,7 @@ import (
 	"errors"
 	"fmt"
 	"log/slog"
+	"net/url"
 	"path/filepath"
 	"strings"
 )
@@ -35,22 +36,25 @@ func Unqualified(n Name) error {
 const MissingPart = "!MISSING!"
 
 const (
-	defaultHost      = "registry.ollama.ai"
-	defaultNamespace = "library"
-	defaultTag       = "latest"
+	defaultHost           = "registry.ollama.ai"
+	defaultNamespace      = "library"
+	defaultTag            = "latest"
+	defaultProtocolScheme = "https"
 )
 
 // DefaultName returns a name with the default values for the host, namespace,
-// and tag parts. The model and digest parts are empty.
+// tag, and protocol scheme parts. The model and digest parts are empty.
 //
 //   - The default host is ("registry.ollama.ai")
 //   - The default namespace is ("library")
 //   - The default tag is ("latest")
+//   - The default protocol scheme is ("https")
 func DefaultName() Name {
 	return Name{
-		Host:      defaultHost,
-		Namespace: defaultNamespace,
-		Tag:       defaultTag,
+		Host:           defaultHost,
+		Namespace:      defaultNamespace,
+		Tag:            defaultTag,
+		ProtocolScheme: defaultProtocolScheme,
 	}
 }
 
@@ -87,10 +91,11 @@ func (k partKind) String() string {
 // It is not guaranteed to be valid. Use [Name.IsValid] to check if the name
 // is valid.
 type Name struct {
-	Host      string
-	Namespace string
-	Model     string
-	Tag       string
+	Host           string
+	Namespace      string
+	Model          string
+	Tag            string
+	ProtocolScheme string
 }
 
 // ParseName parses and assembles a Name from a name string. The
@@ -160,7 +165,9 @@ func ParseNameBare(s string) Name {
 	}
 
 	scheme, host, ok := strings.Cut(s, "://")
-	if !ok {
+	if ok {
+		n.ProtocolScheme = scheme
+	} else {
 		host = scheme
 	}
 	n.Host = host
@@ -189,12 +196,13 @@ func ParseNameFromFilepath(s string) (n Name) {
 	return n
 }
 
-// Merge merges the host, namespace, and tag parts of the two names,
+// Merge merges the host, namespace, tag, and protocol scheme parts of the two names,
 // preferring the non-empty parts of a.
 func Merge(a, b Name) Name {
 	a.Host = cmp.Or(a.Host, b.Host)
 	a.Namespace = cmp.Or(a.Namespace, b.Namespace)
 	a.Tag = cmp.Or(a.Tag, b.Tag)
+	a.ProtocolScheme = cmp.Or(a.ProtocolScheme, b.ProtocolScheme)
 	return a
 }
 
@@ -305,6 +313,23 @@ func (n Name) EqualFold(o Name) bool {
 		strings.EqualFold(n.Tag, o.Tag)
 }
 
+// BaseURL returns the base URL for the registry.
+func (n Name) BaseURL() *url.URL {
+	return &url.URL{
+		Scheme: n.ProtocolScheme,
+		Host:   n.Host,
+	}
+}
+
+// DisplayNamespaceModel returns the namespace and model joined by "/".
+func (n Name) DisplayNamespaceModel() string {
+	var b strings.Builder
+	b.WriteString(n.Namespace)
+	b.WriteByte('/')
+	b.WriteString(n.Model)
+	return b.String()
+}
+
 func isValidLen(kind partKind, s string) bool {
 	switch kind {
 	case kindHost:

types/model/name_test.go

@@ -32,10 +32,11 @@ func TestParseNameParts(t *testing.T) {
 		{
 			in: "scheme://host:port/namespace/model:tag",
 			want: Name{
-				Host:      "host:port",
-				Namespace: "namespace",
-				Model:     "model",
-				Tag:       "tag",
+				Host:           "host:port",
+				Namespace:      "namespace",
+				Model:          "model",
+				Tag:            "tag",
+				ProtocolScheme: "scheme",
 			},
 			wantFilepath: filepath.Join("host:port", "namespace", "model", "tag"),
 		},

x/create/client/create.go

@@ -12,8 +12,8 @@ import (
 	"fmt"
 	"io"
 
+	"github.com/ollama/ollama/manifest"
 	"github.com/ollama/ollama/progress"
-	"github.com/ollama/ollama/server"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/x/create"
 )
@@ -103,7 +103,7 @@ func CreateModel(opts CreateOptions, p *progress.Progress) error {
 // newLayerCreator returns a LayerCreator callback for creating config/JSON layers.
 func newLayerCreator() create.LayerCreator {
 	return func(r io.Reader, mediaType, name string) (create.LayerInfo, error) {
-		layer, err := server.NewLayer(r, mediaType)
+		layer, err := manifest.NewLayer(r, mediaType)
 		if err != nil {
 			return create.LayerInfo{}, err
 		}
@@ -141,13 +141,13 @@ func createQuantizedLayers(r io.Reader, name, dtype string, shape []int32, quant
 	}
 
 	// Create layer for quantized weight
-	weightLayer, err := server.NewLayer(bytes.NewReader(qweightData), server.MediaTypeImageTensor)
+	weightLayer, err := manifest.NewLayer(bytes.NewReader(qweightData), manifest.MediaTypeImageTensor)
 	if err != nil {
 		return nil, err
 	}
 
 	// Create layer for scales
-	scalesLayer, err := server.NewLayer(bytes.NewReader(scalesData), server.MediaTypeImageTensor)
+	scalesLayer, err := manifest.NewLayer(bytes.NewReader(scalesData), manifest.MediaTypeImageTensor)
 	if err != nil {
 		return nil, err
 	}
@@ -169,7 +169,7 @@ func createQuantizedLayers(r io.Reader, name, dtype string, shape []int32, quant
 
 	// Add qbiases layer if present (affine mode)
 	if qbiasData != nil {
-		qbiasLayer, err := server.NewLayer(bytes.NewReader(qbiasData), server.MediaTypeImageTensor)
+		qbiasLayer, err := manifest.NewLayer(bytes.NewReader(qbiasData), manifest.MediaTypeImageTensor)
 		if err != nil {
 			return nil, err
 		}
@@ -186,7 +186,7 @@ func createQuantizedLayers(r io.Reader, name, dtype string, shape []int32, quant
 
 // createUnquantizedLayer creates a single tensor layer without quantization.
 func createUnquantizedLayer(r io.Reader, name string) ([]create.LayerInfo, error) {
-	layer, err := server.NewLayer(r, server.MediaTypeImageTensor)
+	layer, err := manifest.NewLayer(r, manifest.MediaTypeImageTensor)
 	if err != nil {
 		return nil, err
 	}
@@ -221,15 +221,15 @@ func newManifestWriter(opts CreateOptions, capabilities []string) create.Manifes
 		}
 
 		// Create config layer blob
-		configLayer, err := server.NewLayer(bytes.NewReader(configJSON), "application/vnd.docker.container.image.v1+json")
+		configLayer, err := manifest.NewLayer(bytes.NewReader(configJSON), "application/vnd.docker.container.image.v1+json")
 		if err != nil {
 			return fmt.Errorf("failed to create config layer: %w", err)
 		}
 
-		// Convert LayerInfo to server.Layer
-		serverLayers := make([]server.Layer, 0, len(layers))
+		// Convert LayerInfo to manifest.Layer
+		manifestLayers := make([]manifest.Layer, 0, len(layers))
 		for _, l := range layers {
-			serverLayers = append(serverLayers, server.Layer{
+			manifestLayers = append(manifestLayers, manifest.Layer{
 				MediaType: l.MediaType,
 				Digest:    l.Digest,
 				Size:      l.Size,
@@ -243,19 +243,19 @@ func newManifestWriter(opts CreateOptions, capabilities []string) create.Manifes
 			if err != nil {
 				return err
 			}
-			serverLayers = append(serverLayers, modelfileLayers...)
+			manifestLayers = append(manifestLayers, modelfileLayers...)
 		}
 
-		return server.WriteManifest(name, configLayer, serverLayers)
+		return manifest.WriteManifest(name, configLayer, manifestLayers)
 	}
 }
 
 // createModelfileLayers creates layers for template, system, and license from Modelfile config.
-func createModelfileLayers(mf *ModelfileConfig) ([]server.Layer, error) {
-	var layers []server.Layer
+func createModelfileLayers(mf *ModelfileConfig) ([]manifest.Layer, error) {
+	var layers []manifest.Layer
 
 	if mf.Template != "" {
-		layer, err := server.NewLayer(bytes.NewReader([]byte(mf.Template)), "application/vnd.ollama.image.template")
+		layer, err := manifest.NewLayer(bytes.NewReader([]byte(mf.Template)), "application/vnd.ollama.image.template")
 		if err != nil {
 			return nil, fmt.Errorf("failed to create template layer: %w", err)
 		}
@@ -263,7 +263,7 @@ func createModelfileLayers(mf *ModelfileConfig) ([]server.Layer, error) {
 	}
 
 	if mf.System != "" {
-		layer, err := server.NewLayer(bytes.NewReader([]byte(mf.System)), "application/vnd.ollama.image.system")
+		layer, err := manifest.NewLayer(bytes.NewReader([]byte(mf.System)), "application/vnd.ollama.image.system")
 		if err != nil {
 			return nil, fmt.Errorf("failed to create system layer: %w", err)
 		}
@@ -271,7 +271,7 @@ func createModelfileLayers(mf *ModelfileConfig) ([]server.Layer, error) {
 	}
 
 	if mf.License != "" {
-		layer, err := server.NewLayer(bytes.NewReader([]byte(mf.License)), "application/vnd.ollama.image.license")
+		layer, err := manifest.NewLayer(bytes.NewReader([]byte(mf.License)), "application/vnd.ollama.image.license")
 		if err != nil {
 			return nil, fmt.Errorf("failed to create license layer: %w", err)
 		}

x/imagegen/cmd/engine/main.go

@@ -19,6 +19,7 @@ import (
 	"github.com/ollama/ollama/x/imagegen/mlx"
 	"github.com/ollama/ollama/x/imagegen/models/flux2"
 	"github.com/ollama/ollama/x/imagegen/models/gemma3"
+	"github.com/ollama/ollama/x/imagegen/models/glm4_moe_lite"
 	"github.com/ollama/ollama/x/imagegen/models/gpt_oss"
 	"github.com/ollama/ollama/x/imagegen/models/llama"
 	"github.com/ollama/ollama/x/imagegen/models/qwen_image"
@@ -301,6 +302,8 @@ func load(modelPath string) (Model, error) {
 		return gemma3.Load(modelPath)
 	case "gemma3_text":
 		return gemma3.LoadText(modelPath)
+	case "glm4_moe_lite":
+		return glm4_moe_lite.Load(modelPath)
 	default:
 		return llama.Load(modelPath)
 	}

x/imagegen/mlx/mlx.go

@@ -991,6 +991,19 @@ func Concat(a, b *Array, axis int) *Array {
 	return Concatenate([]*Array{a, b}, axis)
 }
 
+// Stack stacks arrays along a new axis (axis 0 by default)
+func Stack(arrays []*Array, axis int) *Array {
+	handles := make([]C.mlx_array, len(arrays))
+	for i, arr := range arrays {
+		handles[i] = arr.c
+	}
+	vec := C.mlx_vector_array_new_data(&handles[0], C.size_t(len(handles)))
+	res := C.mlx_array_new()
+	C.mlx_stack_axis(&res, vec, C.int(axis), C.default_stream())
+	C.mlx_vector_array_free(vec)
+	return newArray(res)
+}
+
 // Slice slices the array
 func Slice(a *Array, start, stop []int32) *Array {
 	n := len(start)

x/imagegen/models/glm4_moe_lite/glm4_moe_lite.go

@@ -0,0 +1,529 @@
+//go:build mlx
+
+// Package glm4_moe_lite provides the GLM4-MoE-Lite implementation for MLX.
+// This model uses Multi-head Latent Attention (MLA) and Mixture of Experts (MoE).
+package glm4_moe_lite
+
+import (
+	"encoding/json"
+	"fmt"
+	"math"
+	"os"
+	"path/filepath"
+
+	"github.com/ollama/ollama/x/imagegen/cache"
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/nn"
+	"github.com/ollama/ollama/x/imagegen/safetensors"
+	"github.com/ollama/ollama/x/imagegen/tokenizer"
+)
+
+// Config holds GLM4-MoE-Lite model configuration
+type Config struct {
+	HiddenSize            int32   `json:"hidden_size"`
+	NumHiddenLayers       int32   `json:"num_hidden_layers"`
+	IntermediateSize      int32   `json:"intermediate_size"`
+	MoEIntermediateSize   int32   `json:"moe_intermediate_size"`
+	NumAttentionHeads     int32   `json:"num_attention_heads"`
+	NumKeyValueHeads      int32   `json:"num_key_value_heads"`
+	VocabSize             int32   `json:"vocab_size"`
+	RMSNormEps            float32 `json:"rms_norm_eps"`
+	RopeTheta             float32 `json:"rope_theta"`
+	MaxPositionEmbeddings int32   `json:"max_position_embeddings"`
+	AttentionBias         bool    `json:"attention_bias"`
+
+	// MLA (Multi-head Latent Attention) parameters
+	QLoraRank      int32 `json:"q_lora_rank"`
+	KVLoraRank     int32 `json:"kv_lora_rank"`
+	QKRopeHeadDim  int32 `json:"qk_rope_head_dim"`
+	QKNopeHeadDim  int32 `json:"qk_nope_head_dim"`
+	VHeadDim       int32 `json:"v_head_dim"`
+
+	// MoE parameters
+	NRoutedExperts      int32   `json:"n_routed_experts"`
+	NSharedExperts      int32   `json:"n_shared_experts"`
+	NumExpertsPerTok    int32   `json:"num_experts_per_tok"`
+	RoutedScalingFactor float32 `json:"routed_scaling_factor"`
+	NormTopKProb        bool    `json:"norm_topk_prob"`
+	FirstKDenseReplace  int32   `json:"first_k_dense_replace"`
+	NGroup              int32   `json:"n_group"`
+	TopKGroup           int32   `json:"topk_group"`
+
+	// Computed fields
+	QHeadDim int32   `json:"-"` // qk_nope_head_dim + qk_rope_head_dim
+	Scale    float32 `json:"-"` // 1/sqrt(QHeadDim)
+}
+
+// MLAAttention implements Multi-head Latent Attention
+type MLAAttention struct {
+	// Low-rank query projections
+	QAProj      *nn.Linear  `weight:"self_attn.q_a_proj"`
+	QALayerNorm *nn.RMSNorm `weight:"self_attn.q_a_layernorm"`
+	QBProj      *nn.Linear  `weight:"self_attn.q_b_proj"`
+
+	// Low-rank KV projections (with shared rope component)
+	KVAProjWithMQA *nn.Linear  `weight:"self_attn.kv_a_proj_with_mqa"`
+	KVALayerNorm   *nn.RMSNorm `weight:"self_attn.kv_a_layernorm"`
+	KVBProj        *nn.Linear  `weight:"self_attn.kv_b_proj"`
+
+	// Output projection
+	OProj *nn.Linear `weight:"self_attn.o_proj"`
+}
+
+// Forward computes MLA attention output
+func (a *MLAAttention) Forward(x *mlx.Array, c cache.Cache, B, L int32, cfg *Config) *mlx.Array {
+	// Query path: q_a_proj -> layernorm -> q_b_proj
+	q := a.QAProj.Forward(x)
+	q = a.QALayerNorm.Forward(q, cfg.RMSNormEps)
+	q = a.QBProj.Forward(q)
+
+	// Reshape Q: [B, L, num_heads * q_head_dim] -> [B, num_heads, L, q_head_dim]
+	q = mlx.Reshape(q, B, L, cfg.NumAttentionHeads, cfg.QHeadDim)
+	q = mlx.Transpose(q, 0, 2, 1, 3)
+
+	// Split Q into nope and rope parts
+	qNope := mlx.Slice(q, []int32{0, 0, 0, 0}, []int32{B, cfg.NumAttentionHeads, L, cfg.QKNopeHeadDim})
+	qPE := mlx.Slice(q, []int32{0, 0, 0, cfg.QKNopeHeadDim}, []int32{B, cfg.NumAttentionHeads, L, cfg.QHeadDim})
+
+	// KV path: kv_a_proj_with_mqa -> split -> layernorm -> kv_b_proj
+	compressedKV := a.KVAProjWithMQA.Forward(x)
+
+	// Split into compressed_kv and k_pe (shared rope component)
+	kvCompressed := mlx.Slice(compressedKV, []int32{0, 0, 0}, []int32{B, L, cfg.KVLoraRank})
+	kPE := mlx.Slice(compressedKV, []int32{0, 0, cfg.KVLoraRank}, []int32{B, L, cfg.KVLoraRank + cfg.QKRopeHeadDim})
+
+	// k_pe is shared across heads (MQA-style): [B, L, rope_dim] -> [B, 1, L, rope_dim]
+	kPE = mlx.Reshape(kPE, B, L, 1, cfg.QKRopeHeadDim)
+	kPE = mlx.Transpose(kPE, 0, 2, 1, 3)
+
+	// Apply layernorm and project KV
+	kvCompressed = a.KVALayerNorm.Forward(kvCompressed, cfg.RMSNormEps)
+	kv := a.KVBProj.Forward(kvCompressed)
+
+	// Reshape KV: [B, L, num_heads * (qk_nope_head_dim + v_head_dim)]
+	kv = mlx.Reshape(kv, B, L, cfg.NumAttentionHeads, cfg.QKNopeHeadDim+cfg.VHeadDim)
+	kv = mlx.Transpose(kv, 0, 2, 1, 3)
+
+	// Split into k_nope and values
+	kNope := mlx.Slice(kv, []int32{0, 0, 0, 0}, []int32{B, cfg.NumAttentionHeads, L, cfg.QKNopeHeadDim})
+	values := mlx.Slice(kv, []int32{0, 0, 0, cfg.QKNopeHeadDim}, []int32{B, cfg.NumAttentionHeads, L, cfg.QKNopeHeadDim + cfg.VHeadDim})
+
+	// Apply RoPE to the rope parts only
+	offset := 0
+	if c != nil {
+		offset = c.Offset()
+	}
+	qPE = mlx.RoPE(qPE, int(cfg.QKRopeHeadDim), true, cfg.RopeTheta, 1.0, offset)
+	kPE = mlx.RoPE(kPE, int(cfg.QKRopeHeadDim), true, cfg.RopeTheta, 1.0, offset)
+
+	// Repeat k_pe across all heads
+	kPE = mlx.Tile(kPE, []int32{1, cfg.NumAttentionHeads, 1, 1})
+
+	// Concatenate nope and rope parts
+	queries := mlx.Concatenate([]*mlx.Array{qNope, qPE}, 3)
+	keys := mlx.Concatenate([]*mlx.Array{kNope, kPE}, 3)
+
+	// Update KV cache
+	if c != nil {
+		keys, values = c.Update(keys, values, int(L))
+	}
+
+	// Scaled dot product attention
+	out := mlx.ScaledDotProductAttention(queries, keys, values, cfg.Scale, L > 1)
+
+	// Reshape back: [B, num_heads, L, v_head_dim] -> [B, L, num_heads * v_head_dim]
+	out = mlx.Reshape(mlx.Transpose(out, 0, 2, 1, 3), B, L, cfg.NumAttentionHeads*cfg.VHeadDim)
+
+	return a.OProj.Forward(out)
+}
+
+// DenseMLP implements the standard SwiGLU MLP for dense layers
+type DenseMLP struct {
+	GateProj *nn.Linear `weight:"mlp.gate_proj"`
+	UpProj   *nn.Linear `weight:"mlp.up_proj"`
+	DownProj *nn.Linear `weight:"mlp.down_proj"`
+}
+
+// Forward applies the SwiGLU MLP
+func (m *DenseMLP) Forward(x *mlx.Array) *mlx.Array {
+	gate := mlx.SiLU(m.GateProj.Forward(x))
+	up := m.UpProj.Forward(x)
+	return m.DownProj.Forward(mlx.Mul(gate, up))
+}
+
+// MoEGate implements the expert gating mechanism
+type MoEGate struct {
+	Weight                 *mlx.Array `weight:"mlp.gate.weight"`
+	EScoreCorrectionBias   *mlx.Array `weight:"mlp.gate.e_score_correction_bias,optional"`
+}
+
+// Forward computes expert selection indices and scores
+func (g *MoEGate) Forward(x *mlx.Array, cfg *Config) (*mlx.Array, *mlx.Array) {
+	// Compute gate logits: x @ weight.T
+	gates := mlx.Linear(x, mlx.Transpose(g.Weight, 1, 0))
+
+	// Sigmoid scoring
+	scores := mlx.Sigmoid(gates)
+	origScores := scores
+
+	// Add correction bias if present
+	if g.EScoreCorrectionBias != nil {
+		scores = mlx.Add(scores, g.EScoreCorrectionBias)
+	}
+
+	// Group-wise expert selection (simplified for n_group=1)
+	// Select top-k experts
+	topK := cfg.NumExpertsPerTok
+	negScores := mlx.Neg(scores)
+	inds := mlx.Argpartition(negScores, int(topK)-1, -1)
+
+	shape := inds.Shape()
+	inds = mlx.Slice(inds, []int32{0, 0, 0}, []int32{shape[0], shape[1], topK})
+
+	// Get scores for selected experts
+	scores = mlx.TakeAlongAxis(origScores, inds, -1)
+
+	// Normalize if configured
+	if topK > 1 && cfg.NormTopKProb {
+		sumScores := mlx.Sum(scores, -1, true)
+		scores = mlx.Div(scores, sumScores)
+	}
+
+	// Apply routing scaling factor
+	scores = mlx.MulScalar(scores, cfg.RoutedScalingFactor)
+
+	return inds, scores
+}
+
+// SwitchMLP implements the MoE expert computation using stacked weights
+// Note: No weight tags - these are populated manually by stacking expert weights
+type SwitchMLP struct {
+	GateWeight *mlx.Array
+	UpWeight   *mlx.Array
+	DownWeight *mlx.Array
+}
+
+// Forward applies the switched expert MLP
+func (s *SwitchMLP) Forward(x *mlx.Array, indices *mlx.Array, cfg *Config) *mlx.Array {
+	shape := x.Shape()
+	B, L := shape[0], shape[1]
+	topK := cfg.NumExpertsPerTok
+
+	// Expand x for expert computation: [B, L, D] -> [B, L, 1, 1, D]
+	xExpanded := mlx.ExpandDims(mlx.ExpandDims(x, -2), -2)
+
+	// Flatten for gather_mm: [B*L, 1, 1, D]
+	xFlat := mlx.Reshape(xExpanded, B*L, 1, 1, cfg.HiddenSize)
+
+	// Flatten indices: [B, L, topK] -> [B*L, topK]
+	idxFlat := mlx.Reshape(indices, B*L, topK)
+
+	// Sort for efficient gather (when we have many tokens)
+	doSort := B*L >= 64
+	var invOrder *mlx.Array
+	n := B * L * topK
+
+	if doSort {
+		idxAll := mlx.Flatten(idxFlat)
+		order := mlx.Argsort(idxAll, 0)
+		invOrder = mlx.Argsort(order, 0)
+		// Reorder x based on sorted indices
+		xFlat = mlx.ExpandDims(mlx.Take(mlx.Squeeze(xFlat, 1), mlx.FloorDivideScalar(order, topK), 0), 1)
+		idxFlat = mlx.Reshape(mlx.Take(idxAll, order, 0), n, 1)
+	}
+
+	// Expert computation using gather_mm
+	// gate: x @ gate_weight.T (indices are on the rhs/weight side)
+	gate := mlx.GatherMM(xFlat, mlx.Transpose(s.GateWeight, 0, 2, 1), nil, idxFlat, doSort)
+	// up: x @ up_weight.T
+	up := mlx.GatherMM(xFlat, mlx.Transpose(s.UpWeight, 0, 2, 1), nil, idxFlat, doSort)
+
+	// SwiGLU activation
+	hidden := mlx.Mul(mlx.SiLU(gate), up)
+
+	// down: hidden @ down_weight.T
+	down := mlx.GatherMM(hidden, mlx.Transpose(s.DownWeight, 0, 2, 1), nil, idxFlat, doSort)
+
+	// Unsort if we sorted
+	if doSort {
+		down = mlx.Reshape(mlx.Take(mlx.Squeeze(mlx.Squeeze(down, 2), 1), invOrder, 0), B*L, topK, cfg.HiddenSize)
+	} else {
+		down = mlx.Squeeze(down, 2)
+	}
+
+	return mlx.Reshape(down, B, L, topK, cfg.HiddenSize)
+}
+
+// SharedExperts implements the shared expert MLP
+type SharedExperts struct {
+	GateProj *nn.Linear `weight:"mlp.shared_experts.gate_proj"`
+	UpProj   *nn.Linear `weight:"mlp.shared_experts.up_proj"`
+	DownProj *nn.Linear `weight:"mlp.shared_experts.down_proj"`
+}
+
+// Forward applies the shared expert MLP
+func (s *SharedExperts) Forward(x *mlx.Array) *mlx.Array {
+	gate := mlx.SiLU(s.GateProj.Forward(x))
+	up := s.UpProj.Forward(x)
+	return s.DownProj.Forward(mlx.Mul(gate, up))
+}
+
+// MoE implements the full Mixture of Experts layer
+type MoE struct {
+	Gate          *MoEGate
+	SwitchMLP     *SwitchMLP
+	SharedExperts *SharedExperts
+}
+
+// Forward applies the MoE layer
+func (m *MoE) Forward(x *mlx.Array, cfg *Config) *mlx.Array {
+	shape := x.Shape()
+	B, L := shape[0], shape[1]
+
+	// Get expert indices and scores
+	inds, scores := m.Gate.Forward(x, cfg)
+
+	// Apply routed experts
+	expertOut := m.SwitchMLP.Forward(x, inds, cfg)
+
+	// Weight by scores: [B, L, topK, D] * [B, L, topK, 1] -> sum over topK
+	scoresExpanded := mlx.ExpandDims(scores, -1)
+	y := mlx.Sum(mlx.Mul(expertOut, scoresExpanded), 2, false)
+
+	// Add shared experts if present
+	if m.SharedExperts != nil {
+		y = mlx.Add(y, m.SharedExperts.Forward(x))
+	}
+
+	return mlx.Reshape(y, B, L, cfg.HiddenSize)
+}
+
+// DenseBlock represents a dense transformer block (for first_k_dense_replace layers)
+type DenseBlock struct {
+	Attention              *MLAAttention
+	MLP                    *DenseMLP
+	InputLayerNorm         *nn.RMSNorm `weight:"input_layernorm"`
+	PostAttentionLayerNorm *nn.RMSNorm `weight:"post_attention_layernorm"`
+}
+
+// Forward applies the dense block
+func (b *DenseBlock) Forward(x *mlx.Array, c cache.Cache, B, L int32, cfg *Config) *mlx.Array {
+	// Pre-norm attention with residual
+	r := b.Attention.Forward(b.InputLayerNorm.Forward(x, cfg.RMSNormEps), c, B, L, cfg)
+	h := mlx.Add(x, r)
+
+	// Pre-norm MLP with residual
+	r = b.MLP.Forward(b.PostAttentionLayerNorm.Forward(h, cfg.RMSNormEps))
+	return mlx.Add(h, r)
+}
+
+// MoEBlock represents a MoE transformer block
+type MoEBlock struct {
+	Attention              *MLAAttention
+	MoE                    *MoE
+	InputLayerNorm         *nn.RMSNorm `weight:"input_layernorm"`
+	PostAttentionLayerNorm *nn.RMSNorm `weight:"post_attention_layernorm"`
+}
+
+// Forward applies the MoE block
+func (b *MoEBlock) Forward(x *mlx.Array, c cache.Cache, B, L int32, cfg *Config) *mlx.Array {
+	// Pre-norm attention with residual
+	r := b.Attention.Forward(b.InputLayerNorm.Forward(x, cfg.RMSNormEps), c, B, L, cfg)
+	h := mlx.Add(x, r)
+
+	// Pre-norm MoE with residual
+	r = b.MoE.Forward(b.PostAttentionLayerNorm.Forward(h, cfg.RMSNormEps), cfg)
+	return mlx.Add(h, r)
+}
+
+// Block interface for both dense and MoE blocks
+type Block interface {
+	Forward(x *mlx.Array, c cache.Cache, B, L int32, cfg *Config) *mlx.Array
+}
+
+// Model represents the complete GLM4-MoE-Lite model
+type Model struct {
+	EmbedTokens *nn.Embedding `weight:"model.embed_tokens"`
+	Layers      []Block       `weight:"-"` // Loaded manually due to different block types
+	Norm        *nn.RMSNorm   `weight:"model.norm"`
+	LMHead      *nn.Linear    `weight:"lm_head"`
+
+	tok *tokenizer.Tokenizer
+	*Config
+}
+
+// sanitizeExpertWeights stacks individual expert weights into a single tensor
+func sanitizeExpertWeights(weights *safetensors.ModelWeights, prefix string, numExperts int32) (*mlx.Array, *mlx.Array, *mlx.Array) {
+	var gateWeights, upWeights, downWeights []*mlx.Array
+
+	for e := int32(0); e < numExperts; e++ {
+		gw, _ := weights.GetTensor(fmt.Sprintf("%s.mlp.experts.%d.gate_proj.weight", prefix, e))
+		uw, _ := weights.GetTensor(fmt.Sprintf("%s.mlp.experts.%d.up_proj.weight", prefix, e))
+		dw, _ := weights.GetTensor(fmt.Sprintf("%s.mlp.experts.%d.down_proj.weight", prefix, e))
+
+		if gw != nil {
+			gateWeights = append(gateWeights, gw)
+		}
+		if uw != nil {
+			upWeights = append(upWeights, uw)
+		}
+		if dw != nil {
+			downWeights = append(downWeights, dw)
+		}
+	}
+
+	var stackedGate, stackedUp, stackedDown *mlx.Array
+	if len(gateWeights) > 0 {
+		stackedGate = mlx.Stack(gateWeights, 0)
+	}
+	if len(upWeights) > 0 {
+		stackedUp = mlx.Stack(upWeights, 0)
+	}
+	if len(downWeights) > 0 {
+		stackedDown = mlx.Stack(downWeights, 0)
+	}
+
+	return stackedGate, stackedUp, stackedDown
+}
+
+// Load loads a GLM4-MoE-Lite model from the given path
+func Load(modelPath string) (*Model, error) {
+	data, err := os.ReadFile(filepath.Join(modelPath, "config.json"))
+	if err != nil {
+		return nil, fmt.Errorf("load config: %w", err)
+	}
+
+	var cfg Config
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return nil, fmt.Errorf("parse config: %w", err)
+	}
+
+	// Compute derived fields
+	cfg.QHeadDim = cfg.QKNopeHeadDim + cfg.QKRopeHeadDim
+	cfg.Scale = float32(1.0 / math.Sqrt(float64(cfg.QHeadDim)))
+
+	weights, err := safetensors.LoadModelWeights(modelPath)
+	if err != nil {
+		return nil, fmt.Errorf("load weights: %w", err)
+	}
+
+	tok, err := tokenizer.Load(filepath.Join(modelPath, "tokenizer.json"))
+	if err != nil {
+		return nil, fmt.Errorf("load tokenizer: %w", err)
+	}
+
+	m := &Model{
+		Layers: make([]Block, cfg.NumHiddenLayers),
+		Config: &cfg,
+		tok:    tok,
+	}
+
+	// Load embedding, norm, and lm_head
+	if err := safetensors.LoadModule(m, weights, ""); err != nil {
+		return nil, err
+	}
+
+	// Load layers manually due to different block types
+	for i := int32(0); i < cfg.NumHiddenLayers; i++ {
+		prefix := fmt.Sprintf("model.layers.%d", i)
+
+		// Load attention (same for both block types)
+		attn := &MLAAttention{}
+		if err := safetensors.LoadModule(attn, weights, prefix); err != nil {
+			return nil, fmt.Errorf("layer %d attention: %w", i, err)
+		}
+
+		if i < cfg.FirstKDenseReplace {
+			// Dense block
+			block := &DenseBlock{Attention: attn}
+			if err := safetensors.LoadModule(block, weights, prefix); err != nil {
+				return nil, fmt.Errorf("layer %d dense: %w", i, err)
+			}
+			m.Layers[i] = block
+		} else {
+			// MoE block
+			block := &MoEBlock{Attention: attn}
+			if err := safetensors.LoadModule(block, weights, prefix); err != nil {
+				return nil, fmt.Errorf("layer %d moe block: %w", i, err)
+			}
+
+			// Stack expert weights
+			gateW, upW, downW := sanitizeExpertWeights(weights, prefix, cfg.NRoutedExperts)
+
+			block.MoE = &MoE{
+				Gate: &MoEGate{},
+				SwitchMLP: &SwitchMLP{
+					GateWeight: gateW,
+					UpWeight:   upW,
+					DownWeight: downW,
+				},
+			}
+
+			// Load gate weights
+			if err := safetensors.LoadModule(block.MoE.Gate, weights, prefix); err != nil {
+				return nil, fmt.Errorf("layer %d gate: %w", i, err)
+			}
+
+			// Load shared experts if present
+			if cfg.NSharedExperts > 0 {
+				block.MoE.SharedExperts = &SharedExperts{}
+				if err := safetensors.LoadModule(block.MoE.SharedExperts, weights, prefix); err != nil {
+					return nil, fmt.Errorf("layer %d shared experts: %w", i, err)
+				}
+			}
+
+			m.Layers[i] = block
+		}
+	}
+
+	mlx.Eval(mlx.Collect(m)...)
+	weights.ReleaseAll()
+
+	return m, nil
+}
+
+// Forward computes the forward pass of the model
+func (m *Model) Forward(tokens *mlx.Array, caches []cache.Cache) *mlx.Array {
+	B, L := tokens.Shape()[0], tokens.Shape()[1]
+
+	h := m.EmbedTokens.Forward(tokens)
+
+	for i, layer := range m.Layers {
+		var c cache.Cache
+		if caches != nil {
+			c = caches[i]
+		}
+		h = layer.Forward(h, c, B, L, m.Config)
+	}
+
+	h = m.Norm.Forward(h, m.RMSNormEps)
+	return m.LMHead.Forward(h)
+}
+
+// Interface methods
+
+// NumLayers returns the number of transformer layers
+func (m *Model) NumLayers() int { return len(m.Layers) }
+
+// MaxContextLength returns the maximum context length
+func (m *Model) MaxContextLength() int32 { return m.MaxPositionEmbeddings }
+
+// VocabSize returns the vocabulary size
+func (m *Model) VocabSize() int32 { return m.Config.VocabSize }
+
+// Tokenizer returns the model's tokenizer
+func (m *Model) Tokenizer() *tokenizer.Tokenizer { return m.tok }
+
+// NewCache creates a new KV cache for the model
+func (m *Model) NewCache(maxSeqLen int32) []cache.Cache {
+	caches := make([]cache.Cache, len(m.Layers))
+	for i := range caches {
+		caches[i] = cache.NewKVCache()
+	}
+	return caches
+}
+
+// FormatPrompt applies the GLM-4 chat template
+func (m *Model) FormatPrompt(prompt string) string {
+	return "[gMASK]<sop><|user|>\n" + prompt + "<|assistant|>\n"
+}

x/server/show.go

@@ -9,7 +9,8 @@ import (
 	"strings"
 
 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/x/imagegen"
+	"github.com/ollama/ollama/manifest"
+	"github.com/ollama/ollama/types/model"
 )
 
 // modelConfig represents the HuggingFace config.json structure
@@ -35,22 +36,22 @@ type modelConfig struct {
 
 // GetSafetensorsLLMInfo extracts model information from safetensors LLM models.
 // It reads the config.json layer and returns a map compatible with GGML's KV format.
-func GetSafetensorsLLMInfo(modelName string) (map[string]any, error) {
-	manifest, err := imagegen.LoadManifest(modelName)
+func GetSafetensorsLLMInfo(name model.Name) (map[string]any, error) {
+	mf, err := manifest.ParseNamedManifest(name)
 	if err != nil {
 		return nil, fmt.Errorf("failed to load manifest: %w", err)
 	}
 
 	var config modelConfig
-	if err := manifest.ReadConfigJSON("config.json", &config); err != nil {
+	if err := mf.ReadConfigJSON("config.json", &config); err != nil {
 		return nil, fmt.Errorf("failed to read config.json: %w", err)
 	}
 
 	// Calculate total tensor bytes from manifest layers
 	var totalBytes int64
 	var tensorCount int64
-	for _, layer := range manifest.Manifest.Layers {
-		if layer.MediaType == "application/vnd.ollama.image.tensor" {
+	for _, layer := range mf.Layers {
+		if layer.MediaType == manifest.MediaTypeImageTensor {
 			totalBytes += layer.Size
 			tensorCount++
 		}
@@ -151,27 +152,30 @@ func buildModelInfo(config modelConfig, totalTensorBytes, tensorCount int64) map
 
 // GetSafetensorsTensorInfo extracts tensor information from safetensors model layers.
 // Each tensor is stored as a minimal safetensors file with an 88-byte header containing metadata.
-func GetSafetensorsTensorInfo(modelName string) ([]api.Tensor, error) {
-	manifest, err := imagegen.LoadManifest(modelName)
+func GetSafetensorsTensorInfo(name model.Name) ([]api.Tensor, error) {
+	mf, err := manifest.ParseNamedManifest(name)
 	if err != nil {
 		return nil, fmt.Errorf("failed to load manifest: %w", err)
 	}
 
-	return getTensorInfoFromManifest(manifest)
+	return getTensorInfoFromManifest(mf)
 }
 
 // getTensorInfoFromManifest extracts tensor info from a manifest.
 // This is separated for testability.
-func getTensorInfoFromManifest(manifest *imagegen.ModelManifest) ([]api.Tensor, error) {
+func getTensorInfoFromManifest(mf *manifest.Manifest) ([]api.Tensor, error) {
 	var tensors []api.Tensor
 
-	for _, layer := range manifest.Manifest.Layers {
-		if layer.MediaType != "application/vnd.ollama.image.tensor" {
+	for _, layer := range mf.Layers {
+		if layer.MediaType != manifest.MediaTypeImageTensor {
 			continue
 		}
 
 		// Read the safetensors header from the blob
-		blobPath := manifest.BlobPath(layer.Digest)
+		blobPath, err := manifest.BlobsPath(layer.Digest)
+		if err != nil {
+			continue
+		}
 		info, err := readSafetensorsHeader(blobPath)
 		if err != nil {
 			// Skip tensors we can't read
@@ -197,15 +201,15 @@ func getTensorInfoFromManifest(manifest *imagegen.ModelManifest) ([]api.Tensor,
 // GetSafetensorsDtype returns the quantization type for a safetensors model.
 // If the model is quantized (has _scale tensors), returns the quantization type (e.g., "FP8").
 // Otherwise returns the torch_dtype from config.json.
-func GetSafetensorsDtype(modelName string) (string, error) {
-	manifest, err := imagegen.LoadManifest(modelName)
+func GetSafetensorsDtype(name model.Name) (string, error) {
+	mf, err := manifest.ParseNamedManifest(name)
 	if err != nil {
 		return "", fmt.Errorf("failed to load manifest: %w", err)
 	}
 
 	// Check if model is quantized by looking for _scale tensors
-	for _, layer := range manifest.Manifest.Layers {
-		if layer.MediaType == "application/vnd.ollama.image.tensor" {
+	for _, layer := range mf.Layers {
+		if layer.MediaType == manifest.MediaTypeImageTensor {
 			if strings.HasSuffix(layer.Name, "_scale") {
 				// Model is quantized - return FP8 (affine quantization)
 				return "FP8", nil
@@ -217,7 +221,7 @@ func GetSafetensorsDtype(modelName string) (string, error) {
 	var cfg struct {
 		TorchDtype string `json:"torch_dtype"`
 	}
-	if err := manifest.ReadConfigJSON("config.json", &cfg); err != nil {
+	if err := mf.ReadConfigJSON("config.json", &cfg); err != nil {
 		return "", fmt.Errorf("failed to read config.json: %w", err)
 	}
 

x/server/show_test.go

@@ -8,7 +8,7 @@ import (
 	"path/filepath"
 	"testing"
 
-	"github.com/ollama/ollama/x/imagegen"
+	"github.com/ollama/ollama/manifest"
 )
 
 func TestBuildModelInfo(t *testing.T) {
@@ -451,8 +451,14 @@ func TestParseSafetensorsHeader_Errors(t *testing.T) {
 }
 
 func TestGetTensorInfoFromManifest(t *testing.T) {
-	// Create a temp directory for blobs
+	// Create a temp directory for blobs and set OLLAMA_MODELS
 	tempDir := t.TempDir()
+	t.Setenv("OLLAMA_MODELS", tempDir)
+
+	blobDir := filepath.Join(tempDir, "blobs")
+	if err := os.MkdirAll(blobDir, 0o755); err != nil {
+		t.Fatalf("failed to create blobs dir: %v", err)
+	}
 
 	// Create test tensor blobs
 	tensors := []struct {
@@ -463,26 +469,26 @@ func TestGetTensorInfoFromManifest(t *testing.T) {
 	}{
 		{
 			name:   "model.embed_tokens.weight",
-			digest: "sha256:abc123",
+			digest: "sha256:abc123abc123abc123abc123abc123abc123abc123abc123abc123abc123abc0",
 			dtype:  "BF16",
 			shape:  []int64{262144, 2560},
 		},
 		{
 			name:   "model.layers.0.self_attn.q_proj.weight",
-			digest: "sha256:def456",
+			digest: "sha256:def456def456def456def456def456def456def456def456def456def456def0",
 			dtype:  "BF16",
 			shape:  []int64{2560, 2560},
 		},
 		{
 			name:   "model.norm.weight",
-			digest: "sha256:ghi789",
+			digest: "sha256:789789789789789789789789789789789789789789789789789789789789abc0",
 			dtype:  "F32",
 			shape:  []int64{2560},
 		},
 	}
 
 	// Create blob files
-	var layers []imagegen.ManifestLayer
+	var layers []manifest.Layer
 	for _, tensor := range tensors {
 		// Create safetensors blob
 		header := map[string]any{
@@ -498,15 +504,17 @@ func TestGetTensorInfoFromManifest(t *testing.T) {
 		binary.Write(&buf, binary.LittleEndian, uint64(len(headerJSON)))
 		buf.Write(headerJSON)
 
-		// Write blob file
-		blobName := "sha256-" + tensor.digest[7:]
-		blobPath := filepath.Join(tempDir, blobName)
+		// Write blob file using the digest format expected by GetBlobsPath
+		blobPath, err := manifest.BlobsPath(tensor.digest)
+		if err != nil {
+			t.Fatalf("failed to get blob path: %v", err)
+		}
 		if err := os.WriteFile(blobPath, buf.Bytes(), 0o644); err != nil {
 			t.Fatalf("failed to write blob: %v", err)
 		}
 
-		layers = append(layers, imagegen.ManifestLayer{
-			MediaType: "application/vnd.ollama.image.tensor",
+		layers = append(layers, manifest.Layer{
+			MediaType: manifest.MediaTypeImageTensor,
 			Digest:    tensor.digest,
 			Size:      int64(buf.Len() + 1000), // header + fake data
 			Name:      tensor.name,
@@ -514,21 +522,20 @@ func TestGetTensorInfoFromManifest(t *testing.T) {
 	}
 
 	// Add a non-tensor layer (should be skipped)
-	layers = append(layers, imagegen.ManifestLayer{
+	layers = append(layers, manifest.Layer{
 		MediaType: "application/vnd.ollama.image.json",
-		Digest:    "sha256:config",
+		Digest:    "sha256:0000000000000000000000000000000000000000000000000000000000000000",
 		Size:      100,
 		Name:      "config.json",
 	})
 
-	manifest := &imagegen.ModelManifest{
-		Manifest: &imagegen.Manifest{
-			Layers: layers,
-		},
-		BlobDir: tempDir,
+	mf := &manifest.Manifest{
+		SchemaVersion: 2,
+		MediaType:     "application/vnd.docker.distribution.manifest.v2+json",
+		Layers:        layers,
 	}
 
-	result, err := getTensorInfoFromManifest(manifest)
+	result, err := getTensorInfoFromManifest(mf)
 	if err != nil {
 		t.Fatalf("getTensorInfoFromManifest() error = %v", err)
 	}