Serve vision models with their projector and original weights on vLLM; group the mirror by version

model serve takes --projector, fetched beside the weights and loaded by
llama.cpp with --mmproj, and --format safetensors with --gpu-class and --tp,
which fetches the weights with their config and tokenizer and runs them on
vLLM. The console offers both, F switching between GGUF and original
weights. The mirror view is one row per version with its files on Enter.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

CHANGELOG.md

@@ -76,6 +76,25 @@ is not an update.
 A model from `server.yaml` is reported but not changed: its source is edited
 there.
 
+### Models of several files: vision projectors, original weights
+
+- A vision model's projector (`mmproj`) is fetched beside its GGUF weights
+  and loaded with them: `llmgw ctl model serve REPO --projector FILE`, or the
+  choice offered by the console. The model is marked as taking images. Nodes
+  load it once their agent is this version or later.
+- `--format safetensors` serves a repository's original weights, with their
+  config.json, tokenizer and shard index, on vLLM across `--tp` GPUs of
+  `--gpu-class` (the fleet's one class when there is one). In the console,
+  **F** on Serve a model switches to them.
+- The mirror keeps where it fetched each file beside it, and reports that
+  and the checksum after it restarts. The console's Mirror view shows one
+  row per version of a model, with its files on Enter.
+- An update watches every file of a model, the projector and the tokenizer
+  included; one whose files the mirror cannot trace is found on the
+  registry by name, commit and size, or given with `llmgw ctl model origin`.
+- `llmgw_model_update_available` is published, and `ModelUpdateAvailable`
+  alerts on it.
+
 ### The manual, in the binary
 
 `llmgw docs` reads every document — installation, operation, the

docs/console.md

@@ -111,9 +111,13 @@ which): requests, tokens, errors and spend.
 team or project is guaranteed and may at most hold, and what it holds now.
 **A** adds one; on a row, change or remove it.
 
-**Mirror** is what the model mirror holds, how far each download has got, and
-which models use each file. **A** downloads a URL into it; on a row, where
-nodes fetch it from, or remove it — with a warning when a model still uses it.
+**Mirror** is what the model mirror holds, one row for each version of each
+model — a file, or the set of files a model of several is: shards, a vision
+projector, a config and tokenizer — with how far its download has got and
+which models use it. **A** downloads a URL into it; on a row, its files with
+their checksums and where each was fetched from, the source a model
+definition names for it, or remove it — with a warning when a model still
+uses it.
 
 **Logs** shows a node's agent log, newest at the bottom, errors and warnings
 coloured.
@@ -181,12 +185,18 @@ to restart it. Another control plane keeps its own copy.
   certificate and starts its agent (`sudo llmgw node add`).
 - **Replicate the store** — moves the store onto two more machines under
   Patroni (offered only while it is not replicated).
-- **Serve a model** — Hugging Face's GGUF models: trending, most downloaded,
-  most liked or newest (**S** changes which), or **/** to search. Enter on
-  one lists its files with their sizes; the one chosen is fetched into the
-  mirror, checked, added to the catalog and started, and a replica is asked
-  to say hello (`llmgw ctl model serve`). `mirror.registry` points it at
-  another registry, for a site without the internet.
+- **Serve a model** — Hugging Face's models: trending, most downloaded,
+  most liked or newest (**S** changes which), or **/** to search. **F**
+  switches between GGUF files, which llama.cpp runs on the CPU or a GPU, and
+  original safetensors weights, which vLLM runs on GPUs. Enter on a GGUF
+  repository lists its files with their sizes; a vision model offers its
+  projector too, which is fetched beside it and loaded with it so the model
+  reads images. Enter on original weights offers them with their config and
+  tokenizer, a GPU class and how many GPUs a replica spans. What is chosen is
+  fetched into the mirror, checked, added to the catalog and started, and a
+  replica is asked to say hello (`llmgw ctl model serve`).
+  `mirror.registry` points it at another registry, for a site without the
+  internet.
 - **Upgrade the cluster** to a release: this machine, then every node in turn
   (`sudo llmgw ctl upgrade`).
 - **Restart all replicas**, one at a time.

internal/agent/engines/llamacpp.go

@@ -93,7 +93,7 @@ func (l *LlamaCpp) Commands(s ShardSpec) ([]Command, error) {
 	bin := l.bin()
 	p := s.Profile
 
-	model, err := ggufIn(p.ModelPath)
+	model, projector, err := ggufIn(p.ModelPath)
 	if err != nil {
 		return nil, err
 	}
@@ -120,6 +120,11 @@ func (l *LlamaCpp) Commands(s ShardSpec) ([]Command, error) {
 	if p.ServedModelName != "" {
 		args = append(args, "--alias", p.ServedModelName)
 	}
+	// A vision model's projector, fetched beside it: what turns an image in
+	// a request into something the language model reads.
+	if projector != "" && !hasFlag(p.Args, "--mmproj", "-mm") {
+		args = append(args, "--mmproj", projector)
+	}
 	if !hasFlag(p.Args, "--parallel", "-np") {
 		n := l.cfg.Parallel
 		if n <= 0 {
@@ -203,30 +208,46 @@ func hasFlag(args []string, names ...string) bool {
 //
 // A path that is already a file is returned untouched, which is every
 // single-file model and every hand-staged one.
-func ggufIn(path string) (string, error) {
+//
+// A vision projector in the directory — an mmproj file — is not the model:
+// it is returned second, to be loaded beside it.
+func ggufIn(path string) (model, projector string, err error) {
 	st, err := os.Stat(path)
 	if err != nil || !st.IsDir() {
 		// Not a directory, or not there at all — the second is the engine's to
 		// report, with the message it would have given anyway.
-		return path, nil //nolint:nilerr // a missing path is the engine's error to give
+		return path, "", nil //nolint:nilerr // a missing path is the engine's error to give
 	}
 	entries, err := os.ReadDir(path)
 	if err != nil {
-		return "", err
+		return "", "", err
 	}
-	var names []string
+	var names, projectors []string
 	for _, e := range entries {
-		if !e.IsDir() && strings.HasSuffix(strings.ToLower(e.Name()), ".gguf") {
+		lower := strings.ToLower(e.Name())
+		if e.IsDir() || !strings.HasSuffix(lower, ".gguf") {
+			continue
+		}
+		if strings.Contains(lower, "mmproj") {
+			projectors = append(projectors, e.Name())
+		} else {
 			names = append(names, e.Name())
 		}
 	}
 	if len(names) == 0 {
-		return "", fmt.Errorf("%s holds no .gguf file", path)
+		if len(projectors) > 0 {
+			return "", "", fmt.Errorf("%s holds a vision projector and no model", path)
+		}
+		return "", "", fmt.Errorf("%s holds no .gguf file", path)
 	}
 	// Sorted, so a split model's first part comes first: the parts are named
 	// "-00001-of-00004" and lexical order is shard order.
 	sort.Strings(names)
-	return filepath.Join(path, names[0]), nil
+	if len(projectors) > 0 {
+		sort.Strings(projectors)
+		projector = filepath.Join(path, projectors[0])
+	}
+	return filepath.Join(path, names[0]), projector, nil
 }
 
 // UseInstalled points the engine at a build the agent installed and checked.

internal/agent/engines/llamacpp_test.go

@@ -204,7 +204,7 @@ func TestASetDirectoryResolvesToItsFirstPart(t *testing.T) {
 		}
 	}
 
-	got, err := ggufIn(dir)
+	got, _, err := ggufIn(dir)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -221,7 +221,7 @@ func TestAPlainFilePathIsUntouched(t *testing.T) {
 	if err := os.WriteFile(p, []byte("x"), 0o644); err != nil {
 		t.Fatal(err)
 	}
-	if got, _ := ggufIn(p); got != p {
+	if got, _, _ := ggufIn(p); got != p {
 		t.Errorf("expected the path unchanged, got %s", got)
 	}
 }
@@ -231,7 +231,7 @@ func TestADirectoryWithNoWeightsSaysSo(t *testing.T) {
 	if err := os.WriteFile(filepath.Join(dir, "config.json"), []byte("{}"), 0o644); err != nil {
 		t.Fatal(err)
 	}
-	if _, err := ggufIn(dir); err == nil {
+	if _, _, err := ggufIn(dir); err == nil {
 		t.Error("a directory with no .gguf in it is not a model llama.cpp can load")
 	}
 }
@@ -240,7 +240,7 @@ func TestAMissingPathIsLeftToTheEngine(t *testing.T) {
 	// The engine's own "no such file" names the path and the flag it came
 	// from, which is more use than anything this could say.
 	p := "/nonexistent/model.gguf"
-	got, err := ggufIn(p)
+	got, _, err := ggufIn(p)
 	if err != nil || got != p {
 		t.Errorf("expected it passed through, got %q %v", got, err)
 	}
@@ -361,3 +361,28 @@ func TestLlamaCppSpreadsAReplicaOverItsNodesGPUsByLayer(t *testing.T) {
 		t.Error("a one-GPU replica was given a split mode")
 	}
 }
+
+func TestAVisionProjectorIsLoadedBesideTheModelNotAsIt(t *testing.T) {
+	dir := t.TempDir()
+	// "mmproj" sorts before the model's name; it must still not be taken for it.
+	for _, n := range []string{"gemma-3-4b-it-Q4_K_M.gguf", "mmproj-gemma-3-4b-it-f16.gguf"} {
+		if err := os.WriteFile(filepath.Join(dir, n), []byte("x"), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	cmds, err := (&LlamaCpp{}).Commands(ShardSpec{HTTPPort: 8080, Profile: Profile{ModelPath: dir}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	args := strings.Join(cmds[0].Args, " ")
+	if !strings.Contains(args, "--model "+filepath.Join(dir, "gemma-3-4b-it-Q4_K_M.gguf")) ||
+		!strings.Contains(args, "--mmproj "+filepath.Join(dir, "mmproj-gemma-3-4b-it-f16.gguf")) {
+		t.Errorf("%s", args)
+	}
+	// One the profile names itself is left as it is.
+	cmds, _ = (&LlamaCpp{}).Commands(ShardSpec{HTTPPort: 8080, Profile: Profile{ModelPath: dir,
+		Args: []string{"--mmproj", "/elsewhere.gguf"}}})
+	if strings.Count(strings.Join(cmds[0].Args, " "), "--mmproj") != 1 {
+		t.Errorf("two projectors: %v", cmds[0].Args)
+	}
+}

internal/cli/model_serve.go

@@ -8,6 +8,7 @@ import (
 
 	"github.com/spf13/cobra"
 
+	"gitea.mynodes.xyz/ilker/llmgw/internal/control/adminapi"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/control/modelserve"
 )
 
@@ -24,15 +25,27 @@ func newModelServeCmd(client func() *ctlClient) *cobra.Command {
 			"that the nodes' llama.cpp can load it and sizes its memory from its own\n" +
 			"header, adds it to the catalog, waits for a replica, and asks it to say\n" +
 			"hello. Each step is shown as it happens; the first that fails stops it.\n\n" +
+			"A vision model's projector (--projector) is fetched beside it and loaded with\n" +
+			"it. --format safetensors serves the original weights instead, with their\n" +
+			"config.json and tokenizer, on vLLM across --tp GPUs of --gpu-class.\n\n" +
 			"  llmgw ctl model serve bartowski/Qwen2.5-3B-Instruct-GGUF\n" +
 			"  llmgw ctl model serve Qwen/Qwen2.5-7B-Instruct-GGUF --quant Q5_K_M --name qwen-7b\n" +
-			"  llmgw ctl model serve owner/big-GGUF --on-demand",
+			"  llmgw ctl model serve owner/big-GGUF --on-demand\n" +
+			"  llmgw ctl model serve google/gemma-3-4b-it-GGUF --projector mmproj-model-f16.gguf\n" +
+			"  llmgw ctl model serve Qwen/Qwen2.5-7B-Instruct --format safetensors --gpu-class l40s-48g",
 		Args: cobra.ExactArgs(1),
 		RunE: func(cmd *cobra.Command, args []string) error {
 			req.Repo = args[0]
 			if onDemand {
 				req.Replicas = 0
 			}
+			// One kind of GPU in the fleet: that one, unless told.
+			if req.Format == "safetensors" && req.GPUClass == "" {
+				var sum adminapi.Summary
+				if err := client().get(cmd.Context(), "/api/v1/summary", &sum); err == nil && len(sum.GPUs) == 1 {
+					req.GPUClass = sum.GPUs[0].Class
+				}
+			}
 			body, _ := json.Marshal(req)
 			var job modelserve.Job
 			if err := client().post(cmd.Context(), "/api/v1/models/serve", body, &job); err != nil {
@@ -44,6 +57,10 @@ func newModelServeCmd(client func() *ctlClient) *cobra.Command {
 	f := c.Flags()
 	f.StringVar(&req.Quant, "quant", "", "the quantisation to serve (default Q4_K_M)")
 	f.StringVar(&req.File, "file", "", "the file to serve, instead of choosing by --quant")
+	f.StringVar(&req.Projector, "projector", "", "a vision model's mmproj file, fetched and loaded beside it")
+	f.StringVar(&req.Format, "format", "gguf", "gguf, run by llama.cpp, or safetensors: the original weights, run by vLLM on GPUs")
+	f.StringVar(&req.GPUClass, "gpu-class", "", "for safetensors: the GPUs to run on (default: the fleet's one class)")
+	f.IntVar(&req.TP, "tp", 1, "for safetensors: how many GPUs one replica spans")
 	f.StringVar(&req.Name, "name", "", "the catalog name (default: from the repository)")
 	f.StringSliceVar(&req.Aliases, "alias", nil, "other names clients may use; repeatable")
 	f.StringVar(&req.Description, "description", "", "what it is for, shown to clients in /v1/models")

internal/cli/tui_console.go

@@ -292,7 +292,7 @@ func (m consoleModel) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 	case consoleMirrorMsg:
 		m.mirror = msg
 	case consoleHFMsg:
-		if msg.query == m.hf.query && msg.sort == m.hf.sort {
+		if msg.query == m.hf.query && msg.sort == m.hf.sort && msg.st == m.hf.safetensors {
 			m.hf.models, m.hf.err = msg.models, msg.err
 			if m.hf.models == nil && msg.err == nil {
 				m.hf.models = []hf.Model{}
@@ -447,7 +447,7 @@ func (m consoleModel) key(k tea.KeyMsg) (tea.Model, tea.Cmd) {
 		case cQuotas:
 			m.menuD = m.quotaMenu(m.quotas[m.row])
 		case cMirror:
-			m.menuD = m.mirrorMenu(m.mirror.res.Models[m.row])
+			m.menuD = m.mirrorMenu(m.mirrorGroups()[m.row])
 		case cServe:
 			return m, m.fetchRepo(m.hf.models[m.row].ID)
 		case cAccounts:
@@ -464,6 +464,12 @@ func (m consoleModel) key(k tea.KeyMsg) (tea.Model, tea.Cmd) {
 		if m.menu == cServe {
 			m.hfSearchDialog()
 		}
+	case "f":
+		if m.menu == cServe {
+			m.hf.safetensors = !m.hf.safetensors
+			m.hf.models, m.hf.err, m.row = nil, nil, 0
+			return m, m.fetchHF()
+		}
 	case "s":
 		if m.menu == cServe {
 			m.hf.sort = (m.hf.sort + 1) % len(hfSorts)
@@ -824,7 +830,7 @@ func (m consoleModel) rowCount() int {
 	case cQuotas:
 		return len(m.quotas)
 	case cMirror:
-		return len(m.mirror.res.Models)
+		return len(m.mirrorGroups())
 	case cServe:
 		return len(m.hf.models)
 	case cAccounts:

internal/cli/tui_hf.go

@@ -11,6 +11,7 @@ import (
 	tea "github.com/charmbracelet/bubbletea"
 
 	"gitea.mynodes.xyz/ilker/llmgw/internal/config"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/control/modelserve"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/hf"
 )
 
@@ -24,15 +25,19 @@ var hfSorts = []struct {
 }{{"trending", hf.SortTrending}, {"most downloaded", hf.SortDownloads}, {"most liked", hf.SortLikes}, {"newest", hf.SortCreated}}
 
 type hfState struct {
-	query  string
-	sort   int
-	models []hf.Model
-	err    error
+	query string
+	sort  int
+	// safetensors lists original weights, for vLLM on GPUs, rather than
+	// GGUF files for llama.cpp.
+	safetensors bool
+	models      []hf.Model
+	err         error
 }
 
 type consoleHFMsg struct {
 	query  string
 	sort   int
+	st     bool
 	models []hf.Model
 	err    error
 }
@@ -56,18 +61,21 @@ func (m consoleModel) registry() *hf.Client {
 }
 
 func (m consoleModel) fetchHF() tea.Cmd {
-	c, q, sort := m.registry(), m.hf.query, m.hf.sort
+	c, q, sort, st := m.registry(), m.hf.query, m.hf.sort, m.hf.safetensors
 	return func() tea.Msg {
 		ctx, cancel := context.WithTimeout(context.Background(), 25*time.Second)
 		defer cancel()
-		hq := hf.Query{Search: q, Sort: hfSorts[sort].sort, GGUF: true, Limit: 50}
+		hq := hf.Query{Search: q, Sort: hfSorts[sort].sort, GGUF: !st, Limit: 50}
+		if st {
+			hq.Filters = []string{"safetensors"}
+		}
 		// Browsing: models that chat. A search finds the rest too, an
 		// embedding model included.
 		if q == "" {
 			hq.Task = "text-generation"
 		}
 		got, err := c.Explore(ctx, hq)
-		return consoleHFMsg{query: q, sort: sort, models: got, err: err}
+		return consoleHFMsg{query: q, sort: sort, st: st, models: got, err: err}
 	}
 }
 
@@ -85,6 +93,17 @@ func (m *consoleModel) fetchRepo(id string) tea.Cmd {
 // repoMenu offers a repository's GGUF files, smallest first as listed.
 func (m consoleModel) repoMenu(r hf.Repo) *menuDialog {
 	var items []menuItem
+	if m.hf.safetensors {
+		title := r.ID
+		set, err := modelserve.SafetensorsSet(r)
+		if err != nil {
+			return &menuDialog{title: title, items: []menuItem{{err.Error(), func(*consoleModel) tea.Cmd { return nil }}}}
+		}
+		return &menuDialog{title: title, items: []menuItem{{
+			fmt.Sprintf("Original weights for vLLM: %s in %d files", bytesShort(set.Bytes), len(set.Files)),
+			func(m *consoleModel) tea.Cmd { m.serveVLLMForm(r, set); return nil },
+		}}}
+	}
 	for _, set := range hf.Group(r.Files) {
 		if !strings.HasSuffix(strings.ToLower(set.Files[0].Path), ".gguf") {
 			continue
@@ -124,13 +143,19 @@ func (m *consoleModel) serveForm(r hf.Repo, set hf.FileSet) {
 		note += ", licence " + r.License
 	}
 	note += "\nThe mirror fetches it and checks it, it joins the catalog, and a\nreplica is started and asked to say hello. The output is shown as it goes."
+	fields := []formField{
+		textF("name", hf.ModelName(r.ID), "what clients ask for"),
+		textF("replicas", "1", "0 runs none until it is first asked for"),
+		textF("context length", "4096", "what one request may use"),
+		textF("description", "", "optional: shown to clients in /v1/models"),
+	}
+	// A vision model: its projector, loaded beside it, is what reads images.
+	if proj := modelserve.Projectors(r); len(proj) > 0 {
+		fields = append(fields, choiceF("images", append(proj, "none: text only"), proj[0]))
+		fields[len(fields)-1].hint = "the vision projector, fetched and loaded beside the model"
+	}
 	m.openForm(&formDialog{title: "Serve " + r.ID, note: note,
-		fields: []formField{
-			textF("name", hf.ModelName(r.ID), "what clients ask for"),
-			textF("replicas", "1", "0 runs none until it is first asked for"),
-			textF("context length", "4096", "what one request may use"),
-			textF("description", "", "optional: shown to clients in /v1/models"),
-		},
+		fields: fields,
 		submit: func(m *consoleModel, v []string) (tea.Cmd, error) {
 			n, err := strconv.Atoi(v[1])
 			if err != nil || n < 0 {
@@ -147,6 +172,50 @@ func (m *consoleModel) serveForm(r hf.Repo, set hf.FileSet) {
 			if v[3] != "" {
 				args = append(args, "--description="+v[3])
 			}
+			if len(v) > 4 && !strings.HasPrefix(v[4], "none") {
+				args = append(args, "--projector="+v[4])
+			}
+			return m.runAfter(args...), nil
+		}})
+}
+
+func (m *consoleModel) serveVLLMForm(r hf.Repo, set hf.FileSet) {
+	server := "--server=" + m.client.base
+	var classes []string
+	for _, g := range m.data.sum.GPUs {
+		classes = append(classes, g.Class)
+	}
+	if len(classes) == 0 {
+		m.dialog, m.dialogTitle = "error", "No GPUs"
+		m.dialogText = "Original weights run on vLLM, on GPUs, and this fleet reports none.\n" +
+			"F goes back to GGUF models, which run on the CPU.\n\n" + sDim.Render("any key")
+		return
+	}
+	m.openForm(&formDialog{title: "Serve " + r.ID,
+		note: fmt.Sprintf("%s in %d files: the weights, config.json and the tokenizer.\n"+
+			"Fetched into the mirror and checked, added, and run by vLLM on the GPUs chosen.", bytesShort(set.Bytes), len(set.Files)),
+		fields: []formField{
+			textF("name", hf.ModelName(r.ID), "what clients ask for"),
+			textF("replicas", "1", "0 runs none until it is first asked for"),
+			textF("context length", "8192", "what one request may use; vLLM's --max-model-len"),
+			textF("description", "", "optional: shown to clients in /v1/models"),
+			choiceF("GPU class", classes, classes[0]),
+			textF("GPUs per replica", "1", "tensor parallel: the cards one replica spans"),
+		},
+		submit: func(m *consoleModel, v []string) (tea.Cmd, error) {
+			for i, what := range map[int]string{1: "replicas", 2: "the context length", 5: "GPUs per replica"} {
+				if n, err := strconv.Atoi(v[i]); err != nil || n < 0 || (i == 5 && n < 1) {
+					return nil, fmt.Errorf("%s is a whole number", what)
+				}
+			}
+			args := []string{"ctl", server, "model", "serve", r.ID, "--format=safetensors", "--gpu-class=" + v[4],
+				"--tp=" + v[5], "--name=" + v[0], "--replicas=" + v[1], "--context-length=" + v[2]}
+			if v[1] == "0" {
+				args = append(args, "--on-demand")
+			}
+			if v[3] != "" {
+				args = append(args, "--description="+v[3])
+			}
 			return m.runAfter(args...), nil
 		}})
 }
@@ -164,7 +233,7 @@ func (m *consoleModel) hfSearchDialog() {
 func (m consoleModel) hfPanel(pw int) (string, string, string) {
 	keys := "→ or Enter open · ↑↓ choose · Q leave"
 	if m.inPanel {
-		keys = "↑↓ row · Enter choose a file · / search · S order · ← back · Q leave"
+		keys = "↑↓ row · Enter choose · / search · S order · F GGUF or original weights · ← back · Q leave"
 	}
 	what := hfSorts[m.hf.sort].label
 	if m.hf.query != "" {
@@ -173,7 +242,11 @@ func (m consoleModel) hfPanel(pw int) (string, string, string) {
 	if m.hf.query == "" {
 		what = "text generation, " + what
 	}
-	head := "\n  GGUF models on Hugging Face: " + what + "\n\n"
+	kind := "GGUF models, for llama.cpp"
+	if m.hf.safetensors {
+		kind = "Original weights, for vLLM on GPUs"
+	}
+	head := "\n  " + kind + " on Hugging Face: " + what + "\n\n"
 	switch {
 	case m.hf.err != nil:
 		return "Serve from Hugging Face", head + "  " + sBad.Render(m.hf.err.Error()) +

internal/cli/tui_mirror.go

@@ -62,26 +62,75 @@ func (m *consoleModel) mirrorCall(title string, do func(context.Context, *mirror
 	}
 }
 
-// usedBy names the catalog's models whose source is this file.
-func (m consoleModel) usedBy(base string, s mirror.Status) []string {
-	u := mirror.Entry{Model: s.Model, Version: s.Version, File: s.File}.URL(base)
-	dir := strings.TrimSuffix(mirror.Entry{Model: s.Model, Version: s.Version}.URL(base), "/") + "/"
+// mirrorGroup is one version of one model on the mirror: a file, or the set
+// of files a model of several is fetched as.
+type mirrorGroup struct {
+	model, version string
+	files          []mirror.Status
+}
+
+func (g mirrorGroup) bytes() (done, total int64) {
+	for _, f := range g.files {
+		done += f.BytesDone
+		total += max(f.BytesTotal, f.BytesDone)
+	}
+	return done, total
+}
+
+// source is what a model's definition names for it: the file, or the
+// directory of a set.
+func (g mirrorGroup) source(base string) string {
+	if len(g.files) == 1 {
+		return mirror.Entry{Model: g.model, Version: g.version, File: g.files[0].File}.URL(base)
+	}
+	return strings.TrimSuffix(mirror.Entry{Model: g.model, Version: g.version}.URL(base), "/") + "/"
+}
+
+func (m consoleModel) mirrorGroups() []mirrorGroup {
+	var out []mirrorGroup
+	at := map[string]int{}
+	for _, st := range m.mirror.res.Models {
+		k := st.Model + "/" + st.Version
+		i, ok := at[k]
+		if !ok {
+			i = len(out)
+			at[k] = i
+			out = append(out, mirrorGroup{model: st.Model, version: st.Version})
+		}
+		out[i].files = append(out[i].files, st)
+	}
+	return out
+}
+
+// usedBy names the catalog's models that run from a version's files.
+func (m consoleModel) usedBy(base string, g mirrorGroup) []string {
+	dir := strings.TrimSuffix(mirror.Entry{Model: g.model, Version: g.version}.URL(base), "/") + "/"
 	var names []string
 	for _, md := range m.data.models {
-		if md.Source == u || strings.HasPrefix(md.Source, dir) {
+		if strings.HasPrefix(md.Source, dir) {
 			names = append(names, md.Name)
 		}
 	}
 	return names
 }
 
-func (m consoleModel) mirrorMenu(s mirror.Status) *menuDialog {
-	what := s.Model + "/" + s.Version
-	used := m.usedBy(m.mirror.res.Base, s)
-	items := []menuItem{{"Where nodes fetch it", func(m *consoleModel) tea.Cmd {
+func (m consoleModel) mirrorMenu(g mirrorGroup) *menuDialog {
+	what := g.model + "/" + g.version
+	used := m.usedBy(m.mirror.res.Base, g)
+	items := []menuItem{{"Its files, and where nodes fetch them", func(m *consoleModel) tea.Cmd {
+		var b strings.Builder
+		for _, f := range g.files {
+			sum := ""
+			if f.SHA256 != "" {
+				sum = sDim.Render("  sha256 " + f.SHA256[:min(12, len(f.SHA256))])
+			}
+			fmt.Fprintf(&b, "%-50s %9s%s\n", truncate(f.File, 50), bytesShort(max(f.BytesTotal, f.BytesDone)), sum)
+			if f.Source != "" {
+				b.WriteString(sDim.Render("  from "+truncate(f.Source, 90)) + "\n")
+			}
+		}
 		m.dialog, m.dialogTitle = "error", what
-		m.dialogText = mirror.Entry{Model: s.Model, Version: s.Version, File: s.File}.URL(m.mirror.res.Base) +
-			"\n\n" + sDim.Render("the source to give a model definition · any key")
+		m.dialogText = b.String() + "\nA model's source: " + g.source(m.mirror.res.Base) + "\n\n" + sDim.Render("any key")
 		return nil
 	}}}
 	items = append(items, menuItem{"Remove from the mirror…", func(m *consoleModel) tea.Cmd {
@@ -93,7 +142,7 @@ func (m consoleModel) mirrorMenu(s mirror.Status) *menuDialog {
 		m.confirmD = &confirmDialog{title: "Remove " + what, text: text,
 			yes: func(m *consoleModel) tea.Cmd {
 				return m.mirrorCall("Remove "+what, func(ctx context.Context, c *mirror.Client) (string, error) {
-					removed, err := c.Remove(ctx, s.Model, s.Version)
+					removed, err := c.Remove(ctx, g.model, g.version)
 					return fmt.Sprintf("Removed %d file(s).", len(removed)), err
 				})
 			}}
@@ -139,24 +188,32 @@ func (m consoleModel) mirrorPanel(pw int) (string, string, string) {
 	}
 	head := "\n  " + sDim.Render("Nodes fetch from "+m.mirror.res.Base) + "\n\n"
 	var rows [][]string
-	for _, s := range m.mirror.res.Models {
-		state := string(s.State)
-		switch s.State {
-		case mirror.Fetching:
-			if p, ok := s.Percent(); ok {
-				state = sWarn.Render(fmt.Sprintf("fetching %d%%", p))
-			} else {
-				state = sWarn.Render("fetching " + bytesShort(s.BytesDone))
+	for _, g := range m.mirrorGroups() {
+		done, total := g.bytes()
+		state := "present"
+		for _, f := range g.files {
+			switch f.State {
+			case mirror.Fetching:
+				state = sWarn.Render("fetching")
+				if total > 0 {
+					state = sWarn.Render(fmt.Sprintf("fetching %d%%", done*100/total))
+				}
+			case mirror.Failed:
+				state = sBad.Render("failed")
+			}
+			if f.State != mirror.Present {
+				break
 			}
-		case mirror.Failed:
-			state = sBad.Render("failed")
 		}
-		size := bytesShort(s.BytesDone)
-		used := strings.Join(m.usedBy(m.mirror.res.Base, s), ", ")
+		what := g.files[0].File
+		if len(g.files) > 1 {
+			what = fmt.Sprintf("%d files", len(g.files))
+		}
+		used := strings.Join(m.usedBy(m.mirror.res.Base, g), ", ")
 		if used == "" {
 			used = sDim.Render("unused")
 		}
-		rows = append(rows, []string{s.Model, s.Version, s.File, state, size, used})
+		rows = append(rows, []string{g.model, g.version, what, state, bytesShort(total), used})
 	}
 	if len(rows) == 0 {
 		return "Mirror", head + "  " + sDim.Render("Empty. A downloads one."), keys

internal/control/modelserve/modelserve.go

@@ -87,6 +87,15 @@ type Request struct {
 	Quant string `json:"quant,omitempty"`
 	// File picks it by name instead.
 	File string `json:"file,omitempty"`
+	// Projector is a vision model's mmproj file in the repository, fetched
+	// beside the weights and loaded with them; empty serves text alone.
+	Projector string `json:"projector,omitempty"`
+	// Format is "gguf", run by llama.cpp, or "safetensors": the original
+	// weights with their config and tokenizer, run by vLLM on GPUs of
+	// GPUClass, across TP of them.
+	Format   string `json:"format,omitempty"`
+	GPUClass string `json:"gpu_class,omitempty"`
+	TP       int    `json:"tp,omitempty"`
 	// Name is the catalog name; the repository suggests one otherwise.
 	Name        string   `json:"name,omitempty"`
 	Aliases     []string `json:"aliases,omitempty"`
@@ -179,6 +188,20 @@ func (s *Server) Start(req Request) (Job, error) {
 	if req.Replicas < 0 {
 		return Job{}, errors.New("replicas cannot be negative")
 	}
+	switch req.Format {
+	case "", "gguf":
+		req.Format = "gguf"
+	case "safetensors":
+		if req.GPUClass == "" {
+			return Job{}, errors.New("safetensors weights run on vLLM, on GPUs: name the GPU class, as `llmgw ctl status` lists it")
+		}
+		if req.Projector != "" {
+			return Job{}, errors.New("a projector is for a GGUF model; vLLM reads a vision model's own files")
+		}
+		req.TP = max(req.TP, 1)
+	default:
+		return Job{}, fmt.Errorf("the format is gguf or safetensors, not %q", req.Format)
+	}
 	if req.Quant == "" && req.File == "" {
 		req.Quant = "Q4_K_M"
 	}
@@ -251,11 +274,28 @@ func (s *Server) run(j *Job) {
 		s.fail(j, "registry", fmt.Errorf("%s is gated: it needs an account's approval the mirror does not have", req.Repo))
 		return
 	}
-	set, err := pick(repo, req)
+	pickSet := pick
+	if req.Format == "safetensors" {
+		pickSet = pickSafetensors
+	}
+	set, err := pickSet(repo, req)
 	if err != nil {
 		s.fail(j, "registry", err)
 		return
 	}
+	if req.Projector != "" {
+		i := slices.IndexFunc(repo.Files, func(f hf.File) bool {
+			return (f.Path == req.Projector || mirror.LastPathElement(f.Path) == req.Projector) &&
+				strings.Contains(strings.ToLower(f.Path), "mmproj")
+		})
+		if i < 0 {
+			s.fail(j, "registry", fmt.Errorf("%s has no vision projector %q; it has: %s", repo.ID, req.Projector,
+				strings.Join(Projectors(repo), ", ")))
+			return
+		}
+		set.Files = append(set.Files, repo.Files[i])
+		set.Bytes += repo.Files[i].Bytes
+	}
 	name := req.Name
 	if name == "" {
 		name = hf.ModelName(repo.ID)
@@ -274,10 +314,20 @@ func (s *Server) run(j *Job) {
 	}
 
 	// 3. Whether the nodes' llama.cpp loads it, and how much memory it needs.
+	if req.Format == "safetensors" {
+		// vLLM reads the model's shape from its config.json, which came
+		// with the weights; the GPUs it needs are the class and width asked.
+		s.step(j, "check", Done, fmt.Sprintf("%.1f GB of weights for vLLM on %d × %s", gb(set.Bytes), req.TP, req.GPUClass))
+		s.addAndStart(ctx, j, repo, set, name, source, catalog.Profile{
+			Name: "gpu", Engine: "vllm", GPUClass: req.GPUClass, TP: req.TP, Weight: 1, Priority: 100,
+			Args: []string{"--max-model-len", strconv.Itoa(req.ContextLength)},
+		})
+		return
+	}
 	s.step(j, "check", Running, "reading its header from the mirror")
 	var header gguf.Header
-	if !strings.HasSuffix(source, "/") {
-		v := gguf.Check(ctx, source, s.deps.LlamaVersion)
+	if u := headerURL(source, set); u != "" {
+		v := gguf.Check(ctx, u, s.deps.LlamaVersion)
 		if !v.Servable {
 			s.fail(j, "check", fmt.Errorf("llama.cpp %s on the nodes cannot run this: %s", s.deps.LlamaVersion, v.Reason))
 			return
@@ -302,16 +352,24 @@ func (s *Server) run(j *Job) {
 		}
 	}
 
+	s.addAndStart(ctx, j, repo, set, name, source, catalog.Profile{
+		Name: "cpu", Engine: "llamacpp", Weight: 1, Priority: 100,
+		Args: []string{"--ctx-size", strconv.Itoa(req.CtxPool)}, MinMemoryGB: memGB, CPUCores: cores,
+	})
+}
+
+// addAndStart puts the model in the catalog with its profile, waits for a
+// replica and asks it to answer.
+func (s *Server) addAndStart(ctx context.Context, j *Job, repo hf.Repo, set hf.FileSet, name, source string, prof catalog.Profile) {
+	req := j.Request
 	// 4. The catalog.
 	m := catalog.Model{
 		Name: name, Aliases: req.Aliases, Description: req.Description, Source: source,
 		RoutingProfile: "chat", ContextLength: req.ContextLength, MaxOutput: 1024,
-		Profiles: []catalog.Profile{{
-			Name: "cpu", Engine: "llamacpp", Weight: 1, Priority: 100,
-			Args: []string{"--ctx-size", strconv.Itoa(req.CtxPool)}, MinMemoryGB: memGB, CPUCores: cores,
-		}},
-		Scaling: catalog.Scaling{Min: req.Replicas, Max: max(req.Replicas, 1)},
-		Origin:  &catalog.Origin{Repo: repo.ID, Revision: repo.Revision, Files: originFiles(set)},
+		Profiles:     []catalog.Profile{prof},
+		Scaling:      catalog.Scaling{Min: req.Replicas, Max: max(req.Replicas, 1)},
+		Capabilities: catalog.Capabilities{Vision: req.Projector != ""},
+		Origin:       &catalog.Origin{Repo: repo.ID, Revision: repo.Revision, Files: originFiles(set)},
 	}
 	s.step(j, "catalog", Running, "adding "+name)
 	if err := s.deps.Catalog.Add(ctx, m, req.By); err != nil {
@@ -593,3 +651,72 @@ func orUnknown(arch string) string {
 	}
 	return arch
 }
+
+// Projectors are a repository's vision projectors: the mmproj files a vision
+// model loads beside its weights.
+func Projectors(repo hf.Repo) []string {
+	var out []string
+	for _, f := range repo.Files {
+		if strings.HasSuffix(strings.ToLower(f.Path), ".gguf") && strings.Contains(strings.ToLower(f.Path), "mmproj") {
+			out = append(out, f.Path)
+		}
+	}
+	return out
+}
+
+// headerURL is the file whose header says what the model is: the source
+// itself, or in a directory the first part of the weights, not a projector.
+func headerURL(source string, set hf.FileSet) string {
+	if !strings.HasSuffix(source, "/") {
+		return source
+	}
+	var names []string
+	for _, f := range set.Files {
+		name := mirror.LastPathElement(f.Path)
+		lower := strings.ToLower(name)
+		if strings.HasSuffix(lower, ".gguf") && !strings.Contains(lower, "mmproj") {
+			names = append(names, name)
+		}
+	}
+	if len(names) == 0 {
+		return ""
+	}
+	slices.Sort(names)
+	return source + names[0]
+}
+
+// SafetensorsSet is what serving a repository's original weights fetches.
+func SafetensorsSet(repo hf.Repo) (hf.FileSet, error) { return pickSafetensors(repo, Request{}) }
+
+// pickSafetensors chooses a repository's original weights and the files an
+// engine reads beside them: config.json, the tokenizer, the shard index.
+func pickSafetensors(repo hf.Repo, _ Request) (hf.FileSet, error) {
+	var weights []hf.File
+	for _, f := range repo.Files {
+		if !strings.Contains(f.Path, "/") && strings.HasSuffix(f.Path, ".safetensors") {
+			weights = append(weights, f)
+		}
+	}
+	sets := hf.Group(weights)
+	if len(sets) == 0 {
+		return hf.FileSet{}, fmt.Errorf("%s has no safetensors weights at its top", repo.ID)
+	}
+	// Sharded "model-00001-of-00004" or a single "model.safetensors"; a
+	// repository with a second, consolidated copy keeps it under another name.
+	set := sets[0]
+	for _, st := range sets {
+		if strings.HasPrefix(st.Name, "model") {
+			set = st
+			break
+		}
+	}
+	companions := hf.Companions(repo.Files)
+	if !slices.ContainsFunc(companions, func(f hf.File) bool { return f.Path == "config.json" }) {
+		return hf.FileSet{}, fmt.Errorf("%s has no config.json, which vLLM reads the model's shape from", repo.ID)
+	}
+	for _, f := range companions {
+		set.Files = append(set.Files, f)
+		set.Bytes += f.Bytes
+	}
+	return set, nil
+}

internal/control/modelserve/modelserve_test.go

@@ -6,6 +6,7 @@ import (
 	"encoding/binary"
 	"net/http"
 	"net/http/httptest"
+	"slices"
 	"strings"
 	"sync"
 	"testing"
@@ -223,3 +224,68 @@ func TestARepositoryThatIsNotOneIsRefused(t *testing.T) {
 		}
 	}
 }
+
+func TestAVisionModelIsServedWithItsProjectorBesideIt(t *testing.T) {
+	files := []hf.File{
+		{Path: "gemma-3-4b-it-Q4_K_M.gguf", Bytes: 2 << 30, SHA256: "aa"},
+		{Path: "mmproj-gemma-3-4b-it-f16.gguf", Bytes: 800 << 20, SHA256: "bb"},
+	}
+	s, m, c, _ := setup(t, files, map[string]string{"gemma-3-4b-it-Q4_K_M.gguf": "aa", "mmproj-gemma-3-4b-it-f16.gguf": "bb"})
+	job, err := s.Start(Request{Repo: "google/gemma-3-4b-it-GGUF", Quant: "Q4_K_M",
+		Projector: "mmproj-gemma-3-4b-it-f16.gguf", Replicas: 1})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if j := wait(t, s, job.ID); j.State != Done {
+		t.Fatalf("%s %+v", j.Error, j.Steps)
+	}
+	if len(m.got) != 2 {
+		t.Fatalf("fetched %+v", m.got)
+	}
+	got := c.added[0]
+	if !strings.HasSuffix(got.Source, "/") || !got.Capabilities.Vision || len(got.Origin.Files) != 2 {
+		t.Errorf("%s vision=%v origin=%+v", got.Source, got.Capabilities.Vision, got.Origin)
+	}
+	// A projector the repository does not have is named with those it has.
+	job, _ = s.Start(Request{Repo: "google/gemma-3-4b-it-GGUF", Quant: "Q4_K_M", Projector: "mmproj-x.gguf", Name: "g2"})
+	if j := wait(t, s, job.ID); j.State != Failed || !strings.Contains(j.Error, "mmproj-gemma-3-4b-it-f16.gguf") {
+		t.Errorf("%+v", j)
+	}
+}
+
+func TestOriginalWeightsAreServedOnVLLMWithTheirConfigAndTokenizer(t *testing.T) {
+	files := []hf.File{
+		{Path: "model-00001-of-00002.safetensors", Bytes: 4 << 30, SHA256: "a1"},
+		{Path: "model-00002-of-00002.safetensors", Bytes: 3 << 30, SHA256: "a2"},
+		{Path: "model.safetensors.index.json", Bytes: 30000},
+		{Path: "config.json", Bytes: 700},
+		{Path: "tokenizer.json", Bytes: 7 << 20},
+		{Path: "tokenizer_config.json", Bytes: 7000},
+		{Path: "README.md", Bytes: 9000},
+		{Path: "original/consolidated.safetensors", Bytes: 7 << 30},
+	}
+	s, m, c, _ := setup(t, files, map[string]string{"model-00001-of-00002.safetensors": "a1", "model-00002-of-00002.safetensors": "a2"})
+	if _, err := s.Start(Request{Repo: "Qwen/Qwen2.5-7B-Instruct", Format: "safetensors"}); err == nil {
+		t.Error("safetensors with no GPU class was taken")
+	}
+	job, err := s.Start(Request{Repo: "Qwen/Qwen2.5-7B-Instruct", Format: "safetensors", GPUClass: "l40s-48g", Replicas: 1})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if j := wait(t, s, job.ID); j.State != Done {
+		t.Fatalf("%s %+v", j.Error, j.Steps)
+	}
+	var got []string
+	for _, r := range m.got {
+		got = append(got, r.File)
+	}
+	if len(got) != 6 || slices.Contains(got, "README.md") || slices.Contains(got, "consolidated.safetensors") {
+		t.Errorf("fetched %v", got)
+	}
+	mod := c.added[0]
+	p := mod.Profiles[0]
+	if !strings.HasSuffix(mod.Source, "/") || p.Engine != "vllm" || p.GPUClass != "l40s-48g" || p.TP != 1 ||
+		!slices.Contains(p.Args, "--max-model-len") {
+		t.Errorf("%s %+v", mod.Source, p)
+	}
+}

internal/control/modelserve/updates.go

@@ -509,9 +509,9 @@ func (s *Server) runUpdate(j *Job) {
 	// 3. Whether the nodes load it, and whether it still fits what is
 	// reserved for it.
 	s.step(j, "check", Running, "reading its header from the mirror")
-	detail := "a set of shards; not read"
-	if !strings.HasSuffix(source, "/") {
-		v := gguf.Check(ctx, source, s.deps.LlamaVersion)
+	detail := "no header to read"
+	if u := headerURL(source, set); u != "" {
+		v := gguf.Check(ctx, u, s.deps.LlamaVersion)
 		if !v.Servable {
 			s.fail(j, "check", fmt.Errorf("llama.cpp %s on the nodes cannot run the new weights: %s", s.deps.LlamaVersion, v.Reason))
 			return

internal/hf/hf.go

@@ -68,6 +68,10 @@ type Model struct {
 	// GGUF is what the registry read from the repository's weights, absent
 	// for one that has none.
 	GGUF *GGUFInfo `json:"gguf"`
+	// Safetensors is the registry's summary of original weights.
+	Safetensors *struct {
+		Total int64 `json:"total"`
+	} `json:"safetensors,omitempty"`
 	// Siblings are the repository's files, when asked for with WithFiles.
 	Siblings []struct {
 		RFilename string `json:"rfilename"`
@@ -84,10 +88,13 @@ type GGUFInfo struct {
 
 // Params is the parameter count, 0 when the registry does not know it.
 func (m Model) Params() int64 {
-	if m.GGUF == nil {
-		return 0
+	switch {
+	case m.GGUF != nil:
+		return m.GGUF.Total
+	case m.Safetensors != nil:
+		return m.Safetensors.Total
 	}
-	return m.GGUF.Total
+	return 0
 }
 
 // Architecture is the GGUF architecture, empty when unknown.
@@ -172,7 +179,7 @@ type Query struct {
 // exploreFields is everything a listing shows. Asking for any field with
 // expand returns only the fields asked for, so all of them are named.
 var exploreFields = []string{"downloads", "likes", "pipeline_tag", "tags", "lastModified",
-	"createdAt", "gated", "private", "trendingScore", "gguf"}
+	"createdAt", "gated", "private", "trendingScore", "gguf", "safetensors"}
 
 // Explore lists repositories: the trending ones, the newest, or a search.
 //