MCP takes pictures of real size, by id too, and waits for a cold model

The game's bridge found three things on the PoC. /mcp held its body to
64 KiB — the edit meant for it had raised the memory notes' limit instead —
so a 1024 px concept sent as base64 failed as "expected one JSON-RPC
request object". The notes are back to 64 KiB, /mcp takes 64 MB, and a
call past it is a 413 saying so.

/v1/files took batch files only, so a picture could not be sent once and
named after. purpose vision and user_data are kept too, as long as job
results, and generate_mesh hands a workflow model the id as it is.

generate_image and review_images failed at once while their model was
cold. They wait out an answer that says when to come back — loading,
another model holding the GPU, a full queue — for as long as the call may,
saying who holds the GPU and about how long; a refusal that says no
Retry-After, a spent budget, is said at once.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

CHANGELOG.md

@@ -28,6 +28,18 @@ APIs change between minors without a deprecation period.
 
 ### Added
 
+- **Pictures uploaded to work from**: `/v1/files` takes `purpose=vision` or
+  `user_data` besides `batch`, kept as long as job results, so a picture is
+  sent once and named by its `file-…` id in the MCP tools and a workflow's
+  inputs.
+- **One ComfyUI for an environment's models**: ComfyUI models in one
+  `comfyui.environment` (two or more) share one replica — the catalog
+  makes it a hidden entry, `comfyui-env-<name>`, on the first member's
+  profiles with the most memory any member needs, carrying every member's
+  weights and custom nodes. Asking for any member starts it; it answers for
+  each with that member's workflow and checkpoint, and the members get no
+  replicas of their own. A picture model and a 3D model on one card stop
+  handing it to each other with a cold start each way.
 - **Pictures, 3D models, vision and workflows as MCP tools** at `/mcp`,
   beside the memory notes, each offered only where the gateway has a model
   for it: `generate_image` (with `concept`, `side` and `texture` styles),
@@ -39,6 +51,12 @@ APIs change between minors without a deprecation period.
   id and download URL. A client that takes a stream gets
   `notifications/progress` while a mesh is made and waits up to ten
   minutes; one that does not, 50 seconds, then the job id.
+- **Setup asks where alerts go**: `llmgw server init --webhook URL`
+  (repeatable) and a **Notification webhooks** field on both setup screens.
+  Slack's, Mattermost's and Rocket.Chat's addresses are sent Slack's text,
+  any other the event as JSON (`slack:` or `json:` in front says which).
+  They are added to the cluster's `notify.webhooks` beside those already
+  there, and each is sent a test once the control plane runs on them.
 
 - **Extended thinking on `/v1/messages`**: `thinking: {type: "enabled"}`
   tells a model that thinks by its chat template to (`enable_thinking`),
@@ -53,6 +71,15 @@ APIs change between minors without a deprecation period.
 
 ### Fixed
 
+- A model whose weights and context pool did not fit its card was added
+  and then failed to make its context on every start, taking the card from
+  others each time (Apertus 70B with a 65536-token pool, 8B with 131072, on
+  24 GB). `ctl model serve` shrinks the pool to what the card holds, saying
+  so, and refuses weights that do not fit at all; the portal's serve form
+  refuses such a pool and says the largest that fits.
+- Removing weights from the mirror that a catalog model still loads is
+  refused, naming the model: removed from the mirror alone, the model stayed
+  in the catalog and its replicas went on loading the node's copy.
 - Jobs ran one at a time, so a 3D job waiting for the card held up another
   model's job that could have run. They run side by side, one per model at
   a time and eight in all, on the leader as before.

docs/admin-api.openapi.json

@@ -9292,6 +9292,10 @@
           "origin": {
             "$ref": "#/components/schemas/catalog.Origin",
             "description": "Origin is the registry repository the weights came from, at which commit, and each file's checksum there. Written when a model is served from the registry; for one that was not, worked out from the mirror's record of where it fetched the file. It is what lets the control plane say that the repository has published different weights since, and move the model to them."
+          },
+          "shared": {
+            "$ref": "#/components/schemas/catalog.SharedComfy",
+            "description": "Shared is set on an environment's hidden entry only: the members its replica answers for. Made by the catalog, never written."
           }
         }
       },
@@ -9545,6 +9549,51 @@
           }
         }
       },
+      "catalog.SharedComfy": {
+        "type": "object",
+        "description": "SharedComfy is what an environment's entry carries of its members.",
+        "properties": {
+          "environment": {
+            "type": "string"
+          },
+          "members": {
+            "type": "array",
+            "items": {
+              "$ref": "#/components/schemas/catalog.SharedMember"
+            }
+          }
+        }
+      },
+      "catalog.SharedMember": {
+        "type": "object",
+        "description": "SharedMember is one model a shared replica answers for.",
+        "properties": {
+          "name": {
+            "type": "string"
+          },
+          "aliases": {
+            "type": "array",
+            "items": {
+              "type": "string"
+            }
+          },
+          "source": {
+            "type": "string",
+            "description": "Source is the member's weights, which the replica's node fetches."
+          },
+          "hf_repo": {
+            "type": "string",
+            "description": "HFRepo is the repository its node loads the weights by name from."
+          },
+          "workflow": {
+            "type": "string",
+            "description": "Workflow and Checkpoint are what a request for it is sent."
+          },
+          "checkpoint": {
+            "type": "string"
+          }
+        }
+      },
       "eval.Live": {
         "type": "object",
         "description": "Live is a run as it stands: one running, with how many of its cases are done, or one finished.",

docs/clients.md

@@ -556,6 +556,15 @@ over HTTP cannot write to your disk, so an agent that needs the file
 locally — to import it in Blender — fetches it, or goes through a small
 local bridge that does.
 
+A picture can be uploaded once and named by its id after: `POST /v1/files`
+with `purpose=vision` (or `user_data`) keeps it, for as long as job results
+are kept, as a file of yours that `generate_mesh`, `review_images` and a
+workflow's file inputs take as `file-…`. A call is at most 64 MB.
+
+A model that is not loaded is waited for: a picture or a look at one whose
+model is starting, or waiting for another to give up the GPU, waits as a
+mesh does, saying so, rather than failing with "model loading".
+
 A mesh takes from half a minute to a few minutes. A client that takes the
 answer as a stream (`Accept: text/event-stream`, as Claude Code does) is
 answered that way: `notifications/progress` for the call's

docs/install.md

@@ -1827,6 +1827,23 @@ same event about the same thing is sent at most once an hour, so a budget that
 refuses a thousand requests is one message. `llmgw ctl notify test` sends a
 test event to each webhook and says how each answered.
 
+Setup asks for them: **Notification webhooks** on its first page, or
+`--webhook` to `llmgw server init`, as often as there are webhooks:
+
+```
+sudo llmgw server init --admin ops@example.com \
+  --webhook https://hooks.slack.com/services/… \
+  --webhook json:https://ops.example.com/llmgw
+```
+
+An address of Slack's, or one with `/hooks/` in it (Mattermost, Rocket.Chat)
+or ending in `/slack`, is sent Slack's `{"text": …}`; any other the event as
+JSON; `slack:` or `json:` in front says which. They are added to the
+cluster's settings beside any already there — run again, nothing is added
+twice — and once the control plane runs on them, each is sent a test, whose
+answer setup prints. Events to choose and a signing secret are set in the
+settings afterwards (`llmgw ctl config edit`).
+
 ### Tracing
 
 Metrics say that a model took four minutes to start; a trace says where they

internal/agent/agent.go

@@ -6,6 +6,7 @@ import (
 	"crypto/sha256"
 	"crypto/tls"
 	"encoding/hex"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"io"
@@ -1131,6 +1132,31 @@ func (a *Agent) startShard(ctx context.Context, m *llmgwv1.StartShard, out chan<
 		stop()
 		spec.Profile.ModelPath = path
 	}
+	// A shared ComfyUI replica's members' weights, each fetched like the
+	// model's, and handed to the gate where they landed.
+	if raw := spec.Profile.Env[engines.SharedModelsEnv]; raw != "" {
+		var ms []map[string]string
+		if err := json.Unmarshal([]byte(raw), &ms); err != nil {
+			report(supervisor.Failed, "the shared replica's models are not readable: "+err.Error(), nil)
+			return
+		}
+		for i, m := range ms {
+			stop := a.watchFetch(m["source"], func(p weights.Progress) {
+				send(&llmgwv1.ShardStatus{State: agentsvc.StateToWire(supervisor.Starting),
+					Fetch: &llmgwv1.FetchProgress{Source: p.Source, BytesDone: p.BytesDone, BytesTotal: p.BytesTotal}})
+			})
+			path, err := a.weights().Ensure(ctx, m["source"])
+			stop()
+			if err != nil {
+				report(supervisor.Failed, fmt.Sprintf("model %s: %v", m["name"], err), nil)
+				return
+			}
+			ms[i]["path"] = path
+		}
+		b, _ := json.Marshal(ms)
+		spec.Profile.Env = maps.Clone(spec.Profile.Env)
+		spec.Profile.Env[engines.SharedModelsEnv] = string(b)
+	}
 	// Adapters are fetched like the model, and before the engine starts: it
 	// loads them at start, and one missing is a replica that answers for the
 	// base model under the adapter's name.

internal/agent/engines/comfyui.go

@@ -129,7 +129,7 @@ func (c *ComfyUI) Commands(s ShardSpec) ([]Command, error) {
 	// so taken before baseEnv drops them. Without the repository the gate
 	// links no Hugging Face cache, and sf3d's node went to the network for
 	// its gated repository and was refused with a 401.
-	for _, k := range []string{comfyNodesEnv, comfyHFRepoEnv} {
+	for _, k := range []string{comfyNodesEnv, comfyHFRepoEnv, SharedModelsEnv} {
 		if v := p.Env[k]; v != "" {
 			env[k] = v
 		}
@@ -144,3 +144,8 @@ const comfyNodesEnv = "LLMGW_COMFYUI_NODES"
 // comfyHFRepoEnv is the repository a custom node loads the weights by name
 // from; catalog.HFRepoEnv and comfy.HFRepoEnv by the same name.
 const comfyHFRepoEnv = "LLMGW_COMFYUI_HF_REPO"
+
+// SharedModelsEnv is a shared replica's members — name, source, and the
+// path the agent fetched it to — as JSON; catalog.SharedModelsEnv and
+// comfy.SharedModelsEnv by the same name.
+const SharedModelsEnv = "LLMGW_COMFYUI_MODELS"

internal/cli/serverinit.go

@@ -73,6 +73,9 @@ type serverInit struct {
 	// patroni are user@host of the machines to keep the store on under
 	// Patroni, replicated; this machine is the third etcd vote for two.
 	patroni []string
+	// webhooks are where the cluster's alerts go: put into its settings
+	// once it runs, and each sent a test.
+	webhooks []string
 	// patroniNodes are the same machines already connected to, which the
 	// setup screen does with the password it was given; when set, patroni
 	// only names them.
@@ -182,6 +185,9 @@ func newServerInitCmd() *cobra.Command {
 		"leave the firewall alone, instead of keeping llmgw's internal ports to the cluster's networks")
 	f.BoolVar(&in.postgres, "postgres", false,
 		"keep the store in PostgreSQL on this machine, installing it and moving an existing SQLite store into it")
+	f.StringArrayVar(&in.webhooks, "webhook", nil,
+		"where alerts go — a model parked, a node lost, a budget spent: an http(s) address, Slack's, Mattermost's "+
+			"and Rocket.Chat's sent Slack's text, any other the event as JSON (slack: or json: in front says which); repeatable")
 	f.StringArrayVar(&in.patroni, "patroni", nil,
 		"user@host of a machine to keep the store on, replicated under Patroni; give two or more. "+
 			"Installs PostgreSQL, Patroni and etcd there over ssh, and moves the existing store in")
@@ -249,6 +255,9 @@ func (in *serverInit) run(ctx context.Context) error {
 			fmt.Fprintf(in.out, "  !!    firewall: %v; llmgw's ports answer every network this machine is on\n", err)
 		}
 	}
+	if err := in.ensureWebhooks(ctx); err != nil {
+		fmt.Fprintf(in.out, "  !!    webhooks: %v\n", err)
+	}
 	if in.judgeWanted() && in.configJudgesWith(guardModelName) {
 		// Not fatal: without its judge the baseline policy still checks for
 		// secrets and injection, and a judged request is served flagged.

internal/cli/serverinit_notify.go

@@ -0,0 +1,185 @@
+package cli
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"net/url"
+	"slices"
+	"strings"
+	"time"
+
+	"go.yaml.in/yaml/v3"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/control/adminapi"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/notify"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/settings"
+)
+
+// Where the cluster says what went wrong: a model parked, a node lost, a
+// budget spent. notify.webhooks existed, and setup never asked for one, so a
+// site learnt of a lost node from its users.
+
+// parseWebhook reads a webhook as setup takes it: an address, with "slack:"
+// or "json:" in front to say what it is sent, else what its address says.
+func parseWebhook(s string) (notify.Webhook, error) {
+	s = strings.TrimSpace(s)
+	format := ""
+	for _, f := range []string{"slack", "json"} {
+		if rest, ok := strings.CutPrefix(s, f+":"); ok && !strings.HasPrefix(rest, "//") {
+			format, s = f, rest
+		}
+	}
+	u, err := url.Parse(s)
+	if err != nil || (u.Scheme != "https" && u.Scheme != "http") || u.Host == "" {
+		return notify.Webhook{}, fmt.Errorf("webhook %q: an http(s) address, with slack: or json: in front to say its format", s)
+	}
+	if format == "" {
+		format = webhookFormat(u)
+	}
+	return notify.Webhook{URL: s, Format: format}, nil
+}
+
+// webhookFormat is what an address takes: Slack's own, and the incoming
+// webhooks that read Slack's {"text": …} — Mattermost's and Rocket.Chat's
+// /hooks/, Discord's …/slack — are sent that; anything else the event as
+// JSON.
+func webhookFormat(u *url.URL) string {
+	switch {
+	case u.Host == "hooks.slack.com",
+		strings.Contains(u.Path, "/hooks/"),
+		strings.HasSuffix(u.Path, "/slack"):
+		return "slack"
+	}
+	return "json"
+}
+
+// parseWebhooks reads a list of them, by commas or as given.
+func parseWebhooks(list []string) ([]notify.Webhook, error) {
+	var out []notify.Webhook
+	for _, item := range list {
+		for _, s := range strings.Split(item, ",") {
+			if strings.TrimSpace(s) == "" {
+				continue
+			}
+			w, err := parseWebhook(s)
+			if err != nil {
+				return nil, err
+			}
+			out = append(out, w)
+		}
+	}
+	return out, nil
+}
+
+// webhookEntry is one of notify.webhooks as the settings document has it.
+type webhookEntry struct {
+	URL        string   `yaml:"url"`
+	Format     string   `yaml:"format,omitempty"`
+	Events     []string `yaml:"events,omitempty,flow"`
+	SecretFile string   `yaml:"secret_file,omitempty"`
+}
+
+// withWebhooks adds these webhooks to a settings document's notify.webhooks,
+// and says whether that changed it: one already there, by address, is left
+// as it is, with whatever events and secret it was given.
+func withWebhooks(doc *settings.Doc, add []notify.Webhook) (bool, error) {
+	path := []string{"notify", "webhooks"}
+	var list []webhookEntry
+	if n := doc.Node(path); n != nil {
+		if err := n.Decode(&list); err != nil {
+			return false, fmt.Errorf("notify.webhooks in the cluster's settings: %w", err)
+		}
+	}
+	changed := false
+	for _, w := range add {
+		if !slices.ContainsFunc(list, func(e webhookEntry) bool { return e.URL == w.URL }) {
+			list = append(list, webhookEntry{URL: w.URL, Format: w.Format})
+			changed = true
+		}
+	}
+	if !changed {
+		return false, nil
+	}
+	b, err := yaml.Marshal(list)
+	if err != nil {
+		return false, err
+	}
+	return true, doc.SetYAML(path, string(b))
+}
+
+// ensureWebhooks puts the webhooks given into the cluster's settings and
+// sends each a test, once the control plane runs on them. Said, not fatal: a
+// cluster without its alerts still serves.
+func (in *serverInit) ensureWebhooks(ctx context.Context) error {
+	if len(in.webhooks) == 0 {
+		return nil
+	}
+	hooks, err := parseWebhooks(in.webhooks)
+	if err != nil {
+		return err
+	}
+	c := in.adminAPI()
+	var cur adminapi.SettingsView
+	if err := c.get(ctx, "/api/v1/settings", &cur); err != nil {
+		return err
+	}
+	doc, err := settings.FromBytes("settings", []byte(cur.Body))
+	if err != nil {
+		return err
+	}
+	changed, err := withWebhooks(doc, hooks)
+	if err != nil {
+		return err
+	}
+	if changed {
+		b, err := doc.Bytes()
+		if err != nil {
+			return err
+		}
+		req, _ := json.Marshal(adminapi.SettingsRequest{Base: cur.Version, Body: string(b), Note: "notification webhooks, from setup"})
+		if err := c.put(ctx, "/api/v1/settings", req, &cur); err != nil {
+			return err
+		}
+		in.done("notification webhooks in the cluster's settings (version %d)", cur.Version)
+	} else {
+		in.kept("notification webhooks (already in the cluster's settings)")
+	}
+
+	// The control planes start again on the change; the test is sent by
+	// one that knows every webhook asked for.
+	var res []notify.Result
+	err = waitFor(ctx, 2*time.Minute, "the control plane to run on its webhooks", func() (bool, error) {
+		res = nil
+		if c.post(ctx, "/api/v1/notify/test", nil, &res) != nil {
+			return false, nil
+		}
+		for _, w := range hooks {
+			// Answered by host: the path of a webhook is its secret.
+			if !slices.ContainsFunc(res, func(r notify.Result) bool { return r.URL == notify.Redact(w.URL) }) {
+				return false, nil
+			}
+		}
+		return true, nil
+	})
+	if err != nil {
+		return err
+	}
+	var failed []string
+	for _, r := range res {
+		if !slices.ContainsFunc(hooks, func(w notify.Webhook) bool { return notify.Redact(w.URL) == r.URL }) {
+			continue
+		}
+		if r.Err != "" {
+			failed = append(failed, r.URL+": "+r.Err)
+			continue
+		}
+		in.done("a test notification reached %s", r.URL)
+	}
+	if len(failed) > 0 {
+		return errors.New("a test notification did not arrive: " + strings.Join(failed, "; ") +
+			" (kept in the settings: 'llmgw ctl notify test' tries again)")
+	}
+	return nil
+}

internal/cli/serverinit_notify_test.go

@@ -0,0 +1,122 @@
+package cli
+
+import (
+	"bytes"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/control/adminapi"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/notify"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/settings"
+)
+
+func TestAWebhooksFormatIsWhatItsAddressTakes(t *testing.T) {
+	for in, want := range map[string]string{
+		"https://hooks.slack.com/services/T1/B2/abc":            "slack",
+		"https://chat.example.com/hooks/xyz":                    "slack", // Mattermost, Rocket.Chat
+		"https://discord.com/api/webhooks/1/abc/slack":          "slack",
+		"https://alerts.example.com/llmgw":                      "json",
+		"json:https://hooks.slack.com/services/T1/B2/abc":       "json",
+		"slack:https://alerts.example.com/llmgw":                "slack",
+		"  https://events.pagerduty.com/integration/x/enqueue ": "json",
+	} {
+		w, err := parseWebhook(in)
+		if err != nil || w.Format != want || strings.HasPrefix(w.URL, "slack:") || strings.HasPrefix(w.URL, "json:") {
+			t.Errorf("%q: %+v %v, want %s", in, w, err, want)
+		}
+	}
+	for _, bad := range []string{"hooks.slack.com/x", "ftp://example.com/x", "slack:", "https://"} {
+		if _, err := parseWebhook(bad); err == nil {
+			t.Errorf("%q was taken", bad)
+		}
+	}
+}
+
+func TestWebhooksAreAddedToTheSettingsBesideTheOnesThere(t *testing.T) {
+	doc, err := settings.FromBytes("settings", []byte("router:\n  retries: 2\nnotify:\n  webhooks:\n"+
+		"    - url: https://alerts.example.com/a\n      events: [node_lost]\n      secret_file: /etc/llmgw/secrets/hook\n"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	hooks, _ := parseWebhooks([]string{"https://alerts.example.com/a, https://hooks.slack.com/services/T/B/c"})
+	changed, err := withWebhooks(doc, hooks)
+	if err != nil || !changed {
+		t.Fatalf("changed %v: %v", changed, err)
+	}
+	b, _ := doc.Bytes()
+	got := string(b)
+	for _, want := range []string{"retries: 2", "url: https://alerts.example.com/a", "events: [node_lost]",
+		"secret_file: /etc/llmgw/secrets/hook", "url: https://hooks.slack.com/services/T/B/c", "format: slack"} {
+		if !strings.Contains(got, want) {
+			t.Errorf("%q not in:\n%s", want, got)
+		}
+	}
+	if strings.Count(got, "alerts.example.com/a") != 1 {
+		t.Errorf("a webhook already there was added again:\n%s", got)
+	}
+	if changed, _ := withWebhooks(doc, hooks); changed {
+		t.Error("adding the same webhooks again changed the settings")
+	}
+}
+
+func TestSetupPutsItsWebhooksInTheSettingsAndTestsThemOnceInForce(t *testing.T) {
+	var mu sync.Mutex
+	body, tests := "router:\n  retries: 2\n", 0
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		mu.Lock()
+		defer mu.Unlock()
+		w.Header().Set("Content-Type", "application/json")
+		switch {
+		case r.Method == http.MethodGet && r.URL.Path == "/api/v1/settings":
+			_ = json.NewEncoder(w).Encode(adminapi.SettingsView{Version: 1, Body: body})
+		case r.Method == http.MethodPut && r.URL.Path == "/api/v1/settings":
+			var req adminapi.SettingsRequest
+			_ = json.NewDecoder(r.Body).Decode(&req)
+			body = req.Body
+			_ = json.NewEncoder(w).Encode(adminapi.SettingsView{Version: 2, Body: body})
+		case r.Method == http.MethodPost && r.URL.Path == "/api/v1/notify/test":
+			tests++
+			// The first answer comes from the control plane before it
+			// started again on the change: it knows no webhooks.
+			res := []notify.Result{}
+			if tests > 1 {
+				res = []notify.Result{{URL: notify.Redact("https://hooks.slack.com/services/T/B/c")},
+					{URL: notify.Redact("https://ops.example.com/in"), Err: "502 Bad Gateway"}}
+			}
+			_ = json.NewEncoder(w).Encode(res)
+		default:
+			http.NotFound(w, r)
+		}
+	}))
+	defer srv.Close()
+	dir := t.TempDir()
+	if err := os.WriteFile(filepath.Join(dir, "server.yaml"),
+		[]byte("admin_listen: "+strings.TrimPrefix(srv.URL, "http://")+"\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	var out bytes.Buffer
+	in := &serverInit{etcDir: dir, out: &out,
+		webhooks: []string{"https://hooks.slack.com/services/T/B/c", "https://ops.example.com/in"}}
+	err := in.ensureWebhooks(t.Context())
+	if err == nil || !strings.Contains(err.Error(), "502") || !strings.Contains(err.Error(), "ctl notify test") {
+		t.Errorf("a webhook that did not answer: %v", err)
+	}
+	mu.Lock()
+	defer mu.Unlock()
+	if !strings.Contains(body, "format: slack") || !strings.Contains(body, "url: https://ops.example.com/in") ||
+		!strings.Contains(body, "retries: 2") || tests < 2 {
+		t.Errorf("settings %q after %d tests", body, tests)
+	}
+	if strings.Contains(out.String(), "/services/T/B/c") {
+		t.Errorf("a webhook's secret path was printed:\n%s", out.String())
+	}
+	if !strings.Contains(out.String(), "reached https://hooks.slack.com/…") {
+		t.Errorf("the one that answered is not said:\n%s", out.String())
+	}
+}

internal/cli/tui_init.go

@@ -50,6 +50,7 @@ func (w *lineWriter) Write(p []byte) (int, error) {
 const (
 	fAdmin = iota
 	fHosts
+	fWebhooks
 	fPostgres
 	fMirror
 	fNode
@@ -70,6 +71,7 @@ type initModel struct {
 	focus     int
 	admin     textField
 	hosts     textField
+	webhooks  textField
 	bundle    textField
 	bundleSHA textField
 	toggles   [fCount]bool
@@ -103,6 +105,11 @@ func newInitModel(in *serverInit, found []string, askForm bool) initModel {
 	hosts.SetValue(strings.Join(in.hosts, ", "))
 	hosts.CharLimit = 400
 	hosts.Width = 40
+	webhooks := newTextField()
+	webhooks.Placeholder = "https://hooks.slack.com/services/…"
+	webhooks.SetValue(strings.Join(in.webhooks, ", "))
+	webhooks.CharLimit = 800
+	webhooks.Width = 40
 	bundle := newTextField()
 	bundle.Placeholder = "/media/usb/llmgw-bundle.tar"
 	bundle.SetValue(in.bundle)
@@ -114,7 +121,7 @@ func newInitModel(in *serverInit, found []string, askForm bool) initModel {
 	bundleSHA.CharLimit = 64
 	bundleSHA.Width = 64
 	sp := newSpinner()
-	m := initModel{in: in, found: found, form: askForm, admin: admin, hosts: hosts,
+	m := initModel{in: in, found: found, form: askForm, admin: admin, hosts: hosts, webhooks: webhooks,
 		bundle: bundle, bundleSHA: bundleSHA, spin: sp, width: 100}
 	m.toggles[fPostgres] = in.postgres
 	m.toggles[fMirror] = !in.noMirror
@@ -235,6 +242,17 @@ func (m initModel) formKey(k tea.KeyMsg) (tea.Model, tea.Cmd) {
 			m.moveFocus(0)
 			return m, nil
 		}
+		hooks := strings.TrimSpace(m.webhooks.Value())
+		if _, err := parseWebhooks([]string{hooks}); err != nil {
+			m.formError = err.Error()
+			m.focus = fWebhooks
+			m.moveFocus(0)
+			return m, nil
+		}
+		m.in.webhooks = nil
+		if hooks != "" {
+			m.in.webhooks = []string{hooks}
+		}
 		m.in.bundle, m.in.bundleSHA = bundle, sum
 		m.in.admin = admin
 		m.in.hosts = nil
@@ -260,6 +278,10 @@ func (m initModel) formKey(k tea.KeyMsg) (tea.Model, tea.Cmd) {
 		var cmd tea.Cmd
 		m.hosts, cmd = m.hosts.Update(k)
 		return m, cmd
+	case fWebhooks:
+		var cmd tea.Cmd
+		m.webhooks, cmd = m.webhooks.Update(k)
+		return m, cmd
 	case fBundle:
 		var cmd tea.Cmd
 		m.bundle, cmd = m.bundle.Update(k)
@@ -284,6 +306,7 @@ func (m *initModel) moveFocus(d int) {
 	}
 	m.admin.Blur()
 	m.hosts.Blur()
+	m.webhooks.Blur()
 	m.bundle.Blur()
 	m.bundleSHA.Blur()
 	switch m.focus {
@@ -291,6 +314,8 @@ func (m *initModel) moveFocus(d int) {
 		m.admin.Focus()
 	case fHosts:
 		m.hosts.Focus()
+	case fWebhooks:
+		m.webhooks.Focus()
 	case fBundle:
 		m.bundle.Focus()
 	case fBundleSHA:
@@ -409,6 +434,8 @@ func (m initModel) formView() string {
 		found = strings.Join(m.found, ", ")
 	}
 	b.WriteString("    " + sDim.Render("Already in the certificate: "+found+". Add a DNS name or NAT address here.") + "\n\n")
+	fmt.Fprintf(&b, "%s\n    %s\n", label(fWebhooks, "Notification webhooks"), m.webhooks.View())
+	b.WriteString("    " + sDim.Render("Where a parked model, a lost node or a spent budget is said: Slack, Mattermost, or any URL taking JSON. Comma-separated; a test is sent.") + "\n\n")
 	b.WriteString(check(fPostgres, "PostgreSQL", "rather than SQLite; installed here, for several control planes or a larger store"))
 	if m.in.mirrorURL != "" {
 		b.WriteString("  Model mirror " + sDim.Render("at "+m.in.mirrorURL+", from --mirror") + "\n")

internal/cli/tui_init_test.go

@@ -20,6 +20,8 @@ func TestTheInitFormBecomesInitsSettings(t *testing.T) {
 	m = typeIn(m, "ops@example.com")
 	m = send(m, tea.KeyMsg{Type: tea.KeyTab})
 	m = typeIn(m, "llm.example.com, 203.0.113.10")
+	m = send(m, tea.KeyMsg{Type: tea.KeyTab})
+	m = typeIn(m, "https://hooks.slack.com/services/T1/B2/x, json:https://alerts.example.com/in")
 	m = send(m, tea.KeyMsg{Type: tea.KeyTab}, tea.KeyMsg{Type: tea.KeyRunes, Runes: []rune(" ")}) // PostgreSQL on
 	m = send(m, tea.KeyMsg{Type: tea.KeyTab}, tea.KeyMsg{Type: tea.KeyTab}, tea.KeyMsg{Type: tea.KeyTab},
 		tea.KeyMsg{Type: tea.KeyRunes, Runes: []rune(" ")}) // starter model off
@@ -33,11 +35,30 @@ func TestTheInitFormBecomesInitsSettings(t *testing.T) {
 	if !in.postgres || in.noMirror || in.noNode || !in.noModel || !in.noGuardModel {
 		t.Errorf("toggles: postgres %v mirror %v node %v model %v", in.postgres, !in.noMirror, !in.noNode, !in.noModel)
 	}
+	if hooks, err := parseWebhooks(in.webhooks); err != nil || len(hooks) != 2 || hooks[0].Format != "slack" || hooks[1].Format != "json" {
+		t.Errorf("webhooks %q: %+v %v", in.webhooks, hooks, err)
+	}
 	if m.(initModel).form {
 		t.Error("the form stayed up after Set up")
 	}
 }
 
+func TestAWebhookThatIsNoAddressKeepsTheFormUp(t *testing.T) {
+	in := &serverInit{}
+	var m tea.Model = newInitModel(in, nil, true)
+	for range fWebhooks {
+		m = send(m, tea.KeyMsg{Type: tea.KeyTab})
+	}
+	m = typeIn(m, "hooks.slack.com/services/x")
+	for range fStart - fWebhooks {
+		m = send(m, tea.KeyMsg{Type: tea.KeyTab})
+	}
+	m = send(m, tea.KeyMsg{Type: tea.KeyEnter})
+	if !m.(initModel).form || m.(initModel).focus != fWebhooks || !strings.Contains(m.View(), "http(s) address") {
+		t.Errorf("an address without its scheme was taken:\n%s", m.View())
+	}
+}
+
 func TestAnOfflineBundleIsGivenInTheFormWithItsDigest(t *testing.T) {
 	in := &serverInit{}
 	var m tea.Model = newInitModel(in, nil, true)
@@ -117,9 +138,10 @@ func TestWithAMirrorElsewhereTheFormDoesNotOfferOne(t *testing.T) {
 	if v := m.View(); !strings.Contains(v, "at http://10.0.0.150:8082, from --mirror") {
 		t.Errorf("the mirror in use is not shown:\n%s", v)
 	}
-	// Tab from the addresses goes past the mirror line to PostgreSQL, then
-	// the node.
-	m = send(m, tea.KeyMsg{Type: tea.KeyTab}, tea.KeyMsg{Type: tea.KeyTab}, tea.KeyMsg{Type: tea.KeyTab})
+	// Tab from the addresses goes to the webhooks, past the mirror line to
+	// PostgreSQL, then the node.
+	m = send(m, tea.KeyMsg{Type: tea.KeyTab}, tea.KeyMsg{Type: tea.KeyTab}, tea.KeyMsg{Type: tea.KeyTab},
+		tea.KeyMsg{Type: tea.KeyTab})
 	if f := m.(initModel).focus; f != fNode {
 		t.Errorf("focus %d, want the node toggle %d", f, fNode)
 	}

internal/cli/tui_setup.go

@@ -15,6 +15,7 @@ import (
 	"github.com/charmbracelet/lipgloss"
 
 	joinpkg "gitea.mynodes.xyz/ilker/llmgw/internal/join"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/notify"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/onboard"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/pgha"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/preflight"
@@ -150,8 +151,8 @@ type setupModel struct {
 	ips    []string
 	choice int // 0 one machine, 1 high availability
 
-	admin, names textField
-	afocus       int // admin, names, next
+	admin, names, webhooks textField
+	afocus                 int // admin, names, webhooks, next
 
 	db     [2]dbMachine
 	mfocus int // over both machines' fields, then the check button
@@ -184,8 +185,10 @@ func newSetupModel(host string, ips []string) setupModel {
 	admin.Placeholder, admin.CharLimit, admin.Width = "you@example.com", 200, 40
 	names := newTextField()
 	names.Placeholder, names.CharLimit, names.Width = "llm.example.com, 203.0.113.10", 400, 40
+	webhooks := newTextField()
+	webhooks.Placeholder, webhooks.CharLimit, webhooks.Width = "https://hooks.slack.com/services/…", 800, 40
 	return setupModel{
-		ref: &progRef{}, host: host, ips: ips, admin: admin, names: names,
+		ref: &progRef{}, host: host, ips: ips, admin: admin, names: names, webhooks: webhooks,
 		db: [2]dbMachine{newDBMachine(1), newDBMachine(2)}, spin: newSpinner(), width: 100, height: 30,
 		log: &[]string{}, logMu: &sync.Mutex{},
 	}
@@ -308,14 +311,17 @@ func (m setupModel) key(k tea.KeyMsg) (tea.Model, tea.Cmd) {
 
 func (m setupModel) adminKey(k tea.KeyMsg) (tea.Model, tea.Cmd) {
 	move := func(d int) {
-		m.afocus = (m.afocus + d + 3) % 3
+		m.afocus = (m.afocus + d + 4) % 4
 		m.admin.Blur()
 		m.names.Blur()
+		m.webhooks.Blur()
 		switch m.afocus {
 		case 0:
 			m.admin.Focus()
 		case 1:
 			m.names.Focus()
+		case 2:
+			m.webhooks.Focus()
 		}
 	}
 	switch k.Type {
@@ -326,7 +332,7 @@ func (m setupModel) adminKey(k tea.KeyMsg) (tea.Model, tea.Cmd) {
 		move(-1)
 		return m, nil
 	case tea.KeyEnter:
-		if m.afocus < 2 {
+		if m.afocus < 3 {
 			move(1)
 			return m, nil
 		}
@@ -336,6 +342,12 @@ func (m setupModel) adminKey(k tea.KeyMsg) (tea.Model, tea.Cmd) {
 			move(0)
 			return m, nil
 		}
+		if _, err := parseWebhooks(m.webhookList()); err != nil {
+			m.formErr = err.Error()
+			m.afocus = 2
+			move(0)
+			return m, nil
+		}
 		m.formErr = ""
 		m.page, m.mfocus = pMachines, 0
 		m.db[0].addr.Focus()
@@ -347,10 +359,23 @@ func (m setupModel) adminKey(k tea.KeyMsg) (tea.Model, tea.Cmd) {
 		m.admin, cmd = m.admin.Update(k)
 	case 1:
 		m.names, cmd = m.names.Update(k)
+	case 2:
+		m.webhooks, cmd = m.webhooks.Update(k)
 	}
 	return m, cmd
 }
 
+// webhookList is the webhooks asked for, one an item.
+func (m setupModel) webhookList() []string {
+	var out []string
+	for _, s := range strings.Split(m.webhooks.Value(), ",") {
+		if s = strings.TrimSpace(s); s != "" {
+			out = append(out, s)
+		}
+	}
+	return out
+}
+
 func (m *setupModel) focusMachine() {
 	for i := range m.db {
 		for f := 0; f < mfCount; f++ {
@@ -589,6 +614,9 @@ func (m setupModel) command() string {
 	for _, t := range m.targets() {
 		parts = append(parts, "--patroni "+t)
 	}
+	for _, w := range m.webhookList() {
+		parts = append(parts, "--webhook "+w)
+	}
 	parts = append(parts, "--no-node", "--no-model")
 	return strings.Join(parts, " \\\n    ")
 }
@@ -600,6 +628,7 @@ func (m *setupModel) startRun() tea.Cmd {
 	in.admin = strings.TrimSpace(m.admin.Value())
 	in.hosts = m.namesList()
 	in.patroni = m.targets()
+	in.webhooks = m.webhookList()
 	// The control plane of a replicated setup is not where models run;
 	// nodes are added once it is up.
 	in.noNode, in.noModel = true, true
@@ -759,7 +788,9 @@ func (m setupModel) adminView() string {
 		sDim.Render("A password is generated and shown once, at the end."))
 	fmt.Fprintf(&b, "%s\n    %s\n    %s\n\n", label(1, "Other names nodes will reach this machine by"), m.names.View(),
 		sDim.Render("Known: "+strings.Join(append([]string{m.host}, m.ips...), ", ")+". Add a DNS name or a floating address."))
-	b.WriteString(button(m.afocus == 2, "Next") + "\n")
+	fmt.Fprintf(&b, "%s\n    %s\n    %s\n\n", label(2, "Notification webhooks"), m.webhooks.View(),
+		sDim.Render("Where a parked model, a lost node or a spent budget is said: Slack, Mattermost, or any URL taking JSON. Optional; a test is sent."))
+	b.WriteString(button(m.afocus == 3, "Next") + "\n")
 	return b.String()
 }
 
@@ -880,6 +911,13 @@ func (m setupModel) planView() string {
 	fmt.Fprintf(&b, "  %s  %s\n      %s\n", lipgloss.NewStyle().Bold(true).Render(m.host), sDim.Render(m.thisAddr()),
 		"control plane and gateway, model mirror, etcd (the third vote)")
 	fmt.Fprintf(&b, "      administrator %s\n", strings.TrimSpace(m.admin.Value()))
+	if hooks, _ := parseWebhooks(m.webhookList()); len(hooks) > 0 {
+		var where []string
+		for _, h := range hooks {
+			where = append(where, notify.Redact(h.URL)+" ("+h.Format+")")
+		}
+		fmt.Fprintf(&b, "      alerts to %s\n", strings.Join(where, ", "))
+	}
 	granted := false
 	for _, d := range m.db {
 		addr := strings.TrimSpace(d.addr.Value())

internal/cli/tui_setup_test.go

@@ -45,7 +45,7 @@ func TestTheSetupAsksForWhatHighAvailabilityNeedsAndRefusesWhatWillNotDo(t *test
 		t.Fatalf("high availability did not lead to the administrator: page %d", s.page)
 	}
 	// No email: refused, with a reason.
-	m = press(m, "tab", "tab", "enter")
+	m = press(m, "tab", "tab", "tab", "enter")
 	if s := m.(setupModel); s.page != pAdmin || !strings.Contains(s.formErr, "email") {
 		t.Fatalf("went on without an administrator: %+v", s.formErr)
 	}
@@ -54,7 +54,7 @@ func TestTheSetupAsksForWhatHighAvailabilityNeedsAndRefusesWhatWillNotDo(t *test
 	if !strings.Contains(m.View(), "Not yet") {
 		t.Error("the refusal was not shown as a dialog")
 	}
-	m = press(m, "x", "a@example.com", "tab", "tab", "enter")
+	m = press(m, "x", "a@example.com", "tab", "tab", "tab", "enter")
 	s := m.(setupModel)
 	if s.page != pMachines {
 		t.Fatalf("page %d: %s", s.page, s.formErr)

internal/comfy/gate.go

@@ -20,6 +20,7 @@ import (
 	"os/signal"
 	"path/filepath"
 	"regexp"
+	"slices"
 	"strconv"
 	"strings"
 	"syscall"
@@ -112,7 +113,15 @@ func RunGate(ctx context.Context, cfg GateConfig) error {
 		return err
 	}
 	models := filepath.Join(state, "models")
-	if err := LinkWeights(cfg.Weights, models); err != nil {
+	shared, err := parseShared(os.Getenv(SharedModelsEnv))
+	if err != nil {
+		return err
+	}
+	all := []string{cfg.Weights}
+	for _, m := range shared {
+		all = append(all, m.Path)
+	}
+	if err := LinkAllWeights(all, models); err != nil {
 		return err
 	}
 	if repo := os.Getenv(HFRepoEnv); repo != "" {
@@ -120,6 +129,20 @@ func RunGate(ctx context.Context, cfg GateConfig) error {
 			return fmt.Errorf("showing the weights as %s: %w", repo, err)
 		}
 	}
+	for _, m := range shared {
+		if m.HFRepo == "" {
+			continue
+		}
+		if err := LinkHFCache(m.Path, m.HFRepo, hfHubDir()); err != nil {
+			return fmt.Errorf("showing %s's weights as %s: %w", m.Name, m.HFRepo, err)
+		}
+	}
+	names := []string{cfg.Model}
+	for _, m := range shared {
+		if !slices.Contains(names, m.Name) {
+			names = append(names, m.Name)
+		}
+	}
 	extra := filepath.Join(state, "extra_model_paths.yaml")
 	if err := os.WriteFile(extra, []byte(extraModelPaths(models)), 0o644); err != nil { //nolint:gosec // not secret
 		return err
@@ -177,7 +200,7 @@ func RunGate(ctx context.Context, cfg GateConfig) error {
 	go book.watch(ctx, upstream, clientID, log)
 	srv := &http.Server{
 		Addr:              cfg.Listen,
-		Handler:           newGate(upstream, cfg.Key, cfg.Model, book, clientID),
+		Handler:           newGate(upstream, cfg.Key, names, book, clientID),
 		ReadHeaderTimeout: 10 * time.Second,
 	}
 	served := make(chan error, 1)
@@ -218,12 +241,12 @@ func shutdown(srv *http.Server) {
 
 // NewGate is the gate's handler, in front of the ComfyUI at upstream.
 func NewGate(upstream *url.URL, key, model string) http.Handler {
-	return newGate(upstream, key, model, newProgressBook(), "")
+	return newGate(upstream, key, []string{model}, newProgressBook(), "")
 }
 
 // newGate is the gate with the book of its workflows' progress, kept under
 // clientID; an empty one leaves the workflows' own client ids alone.
-func newGate(upstream *url.URL, key, model string, book *progressBook, clientID string) http.Handler {
+func newGate(upstream *url.URL, key string, names []string, book *progressBook, clientID string) http.Handler {
 	proxy := httputil.NewSingleHostReverseProxy(upstream)
 	base := proxy.Director
 	proxy.Director = func(r *http.Request) {
@@ -269,7 +292,11 @@ func newGate(upstream *url.URL, key, model string, book *progressBook, clientID
 				return
 			}
 			w.Header().Set("Content-Type", "application/json")
-			fmt.Fprintf(w, `{"object":"list","data":[{"id":%q,"object":"model","owned_by":"comfyui"}]}`, model)
+			data := make([]map[string]string, 0, len(names))
+			for _, n := range names {
+				data = append(data, map[string]string{"id": n, "object": "model", "owned_by": "comfyui"})
+			}
+			_ = json.NewEncoder(w).Encode(map[string]any{"object": "list", "data": data})
 			return
 		}
 		if id, ok := strings.CutPrefix(r.URL.Path, "/progress/"); ok && r.Method == http.MethodGet {
@@ -306,6 +333,106 @@ func FileName(path string) string {
 	return cachedName.ReplaceAllString(filepath.Base(path), "")
 }
 
+// SharedModelsEnv is where the agent hands a shared replica's gate its
+// members: name, the path their weights were fetched to, and the repository
+// a node loads them by name from.
+const SharedModelsEnv = "LLMGW_COMFYUI_MODELS"
+
+// sharedModel is one member of a shared replica.
+type sharedModel struct {
+	Name   string `json:"name"`
+	Path   string `json:"path"`
+	HFRepo string `json:"hf_repo"`
+}
+
+func parseShared(raw string) ([]sharedModel, error) {
+	if strings.TrimSpace(raw) == "" {
+		return nil, nil
+	}
+	var ms []sharedModel
+	if err := json.Unmarshal([]byte(raw), &ms); err != nil {
+		return nil, fmt.Errorf("%s: %w", SharedModelsEnv, err)
+	}
+	for _, m := range ms {
+		if m.Name == "" || m.Path == "" {
+			return nil, fmt.Errorf("%s: a member with no name or no fetched path", SharedModelsEnv)
+		}
+	}
+	return ms, nil
+}
+
+// LinkAllWeights shows ComfyUI every model's weights at once, for a replica
+// that answers for several: links made before are taken away first, then
+// each file of each is linked by its name. Two weights files of one name
+// are refused — ComfyUI would load one for both; anything else of one name
+// (a README, a licence) the first keeps.
+func LinkAllWeights(weights []string, dir string) error {
+	if len(weights) == 1 {
+		return LinkWeights(weights[0], dir)
+	}
+	if err := LinkWeights(weights[0], dir); err != nil {
+		return err
+	}
+	for _, w := range weights[1:] {
+		files, err := weightFiles(w)
+		if err != nil {
+			return err
+		}
+		for _, f := range files {
+			link := filepath.Join(dir, FileName(f))
+			if cur, err := os.Readlink(link); err == nil {
+				if cur == f {
+					continue
+				}
+				if isWeightsFile(f) {
+					return fmt.Errorf("two models' weights are both named %s (%s and %s): give one comfyui.checkpoint of another name",
+						FileName(f), cur, f)
+				}
+				continue
+			}
+			if err := os.Symlink(f, link); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+// weightFiles are the files a weights source is: itself, or each of its
+// directory's.
+func weightFiles(weights string) ([]string, error) {
+	weights, err := filepath.Abs(weights)
+	if err != nil {
+		return nil, err
+	}
+	info, err := os.Stat(weights)
+	if err != nil {
+		return nil, fmt.Errorf("the weights: %w", err)
+	}
+	if !info.IsDir() {
+		return []string{weights}, nil
+	}
+	entries, err := os.ReadDir(weights)
+	if err != nil {
+		return nil, err
+	}
+	var out []string
+	for _, e := range entries {
+		if e.Type().IsRegular() && !strings.HasPrefix(e.Name(), ".") {
+			out = append(out, filepath.Join(weights, e.Name()))
+		}
+	}
+	return out, nil
+}
+
+func isWeightsFile(path string) bool {
+	switch strings.ToLower(filepath.Ext(path)) {
+	case ".safetensors", ".ckpt", ".pt", ".pth", ".bin", ".gguf", ".sft":
+		return true
+	}
+	return false
+}
+
 // LinkWeights shows ComfyUI the weights: a link in dir to the file, or to
 // every file of the directory. Links made before are taken away first, so
 // a replica started again on other weights sees only those.

internal/comfy/gate_test.go

@@ -261,3 +261,41 @@ func TestLLMGWsOwnNodesArePutInEveryReplica(t *testing.T) {
 		t.Fatalf("%v %.80s", err, b)
 	}
 }
+
+func TestASharedReplicaSeesEveryMembersWeights(t *testing.T) {
+	root := t.TempDir()
+	write := func(p string) string {
+		full := filepath.Join(root, p)
+		_ = os.MkdirAll(filepath.Dir(full), 0o755)
+		if err := os.WriteFile(full, []byte("w"), 0o644); err != nil {
+			t.Fatal(err)
+		}
+		return full
+	}
+	flux := write("flux1-schnell-fp8.safetensors")
+	write("sf3d/model.safetensors")
+	write("sf3d/README.md")
+	write("tripo/model.ckpt")
+	write("tripo/README.md")
+	dir := filepath.Join(root, "models")
+	if err := LinkAllWeights([]string{flux, filepath.Join(root, "sf3d"), filepath.Join(root, "tripo")}, dir); err != nil {
+		t.Fatal(err)
+	}
+	for _, want := range []string{"flux1-schnell-fp8.safetensors", "model.safetensors", "model.ckpt", "README.md"} {
+		if _, err := os.Lstat(filepath.Join(dir, want)); err != nil {
+			t.Errorf("%s not linked: %v", want, err)
+		}
+	}
+	// Two weights of one name: ComfyUI would load one for both.
+	write("other/model.safetensors")
+	err := LinkAllWeights([]string{filepath.Join(root, "sf3d"), filepath.Join(root, "other")}, dir)
+	if err == nil || !strings.Contains(err.Error(), "both named model.safetensors") {
+		t.Errorf("err %v", err)
+	}
+	if ms, err := parseShared(`[{"name":"sf3d","path":"/x","hf_repo":"stabilityai/stable-fast-3d"}]`); err != nil || ms[0].HFRepo == "" {
+		t.Errorf("parse %+v %v", ms, err)
+	}
+	if _, err := parseShared(`[{"name":"sf3d"}]`); err == nil {
+		t.Error("a member with no path was taken")
+	}
+}

internal/comfy/progress_test.go

@@ -60,7 +60,7 @@ func TestTheGateHearsTheWorkflowsItLetsThrough(t *testing.T) {
 	defer comfy.Close()
 	u, _ := url.Parse(comfy.URL)
 	book := newProgressBook()
-	gate := httptest.NewServer(newGate(u, "", "m", book, "llmgw-gate-1"))
+	gate := httptest.NewServer(newGate(u, "", []string{"m"}, book, "llmgw-gate-1"))
 	defer gate.Close()
 	resp, err := http.Post(gate.URL+"/prompt", "application/json", strings.NewReader(`{"prompt":{},"client_id":"router"}`))
 	if err != nil {

internal/control/catalog/catalog.go

@@ -409,6 +409,10 @@ type Model struct {
 	// say that the repository has published different weights since, and
 	// move the model to them.
 	Origin *Origin `json:"origin,omitempty" koanf:"origin"`
+
+	// Shared is set on an environment's hidden entry only: the members its
+	// replica answers for. Made by the catalog, never written.
+	Shared *SharedComfy `json:"shared,omitempty" koanf:"-"`
 }
 
 // RoleGuard is a judge model's role.
@@ -650,6 +654,11 @@ const NodesEnv = "LLMGW_COMFYUI_NODES"
 // HFRepoEnv carries the repository a node loads the weights by name from.
 const HFRepoEnv = "LLMGW_COMFYUI_HF_REPO"
 
+// SharedModelsEnv carries a shared replica's members — name, weights source,
+// repository — to the agent, which fetches each and hands the gate where
+// they landed.
+const SharedModelsEnv = "LLMGW_COMFYUI_MODELS"
+
 var hfRepo = regexp.MustCompile(`^[A-Za-z0-9][A-Za-z0-9._-]*/[A-Za-z0-9][A-Za-z0-9._-]*$`)
 
 var (
@@ -778,6 +787,9 @@ func New(models []Model) (*Catalog, error) {
 	}
 
 	for i := range models {
+		if models[i].IsEnvironment() {
+			continue // the catalog's own, made again below
+		}
 		m := models[i].normalize()
 		if err := m.validate(); err != nil {
 			return nil, err
@@ -851,9 +863,25 @@ func New(models []Model) (*Catalog, error) {
 		}
 		canaryOf[target.Name] = name
 	}
-	if _, err := c.environments(); err != nil {
+	envs, err := c.environments()
+	if err != nil {
 		return nil, err
 	}
+	for _, e := range envs {
+		if len(e.Models) < 2 {
+			continue
+		}
+		var members []Model
+		for _, n := range e.Models {
+			members = append(members, c.models[n])
+		}
+		em := envModel(e, members)
+		if _, clash := c.models[em.Name]; clash {
+			return nil, fmt.Errorf("model %q: the name is the hidden entry of environment %q", em.Name, e.Name)
+		}
+		c.models[em.Name] = em
+		c.order = append(c.order, em.Name)
+	}
 	return c, nil
 }
 
@@ -1297,6 +1325,25 @@ func (m Model) EngineProfile(name string) (engines.Profile, bool) {
 			}
 			env[NodesEnv] = string(nodes)
 		}
+		if p.Engine == EngineComfyUI && m.Shared != nil {
+			// Every member's weights, for the node to fetch and the gate to
+			// link: one replica answers for all of them.
+			type member struct {
+				Name   string `json:"name"`
+				Source string `json:"source"`
+				HFRepo string `json:"hf_repo,omitempty"`
+			}
+			var ms []member
+			for _, sm := range m.Shared.Members {
+				ms = append(ms, member{Name: sm.Name, Source: sm.Source, HFRepo: sm.HFRepo})
+			}
+			b, _ := json.Marshal(ms)
+			env = maps.Clone(env)
+			if env == nil {
+				env = map[string]string{}
+			}
+			env[SharedModelsEnv] = string(b)
+		}
 		if p.Engine == EngineComfyUI && m.ComfyUI != nil && m.ComfyUI.HuggingFaceRepo != "" {
 			env = maps.Clone(env)
 			if env == nil {

internal/control/catalog/environments.go

@@ -75,7 +75,7 @@ func (c *Catalog) environments() ([]Environment, error) {
 	for _, name := range c.order {
 		m := c.models[name]
 		env := m.ComfyEnvironment()
-		if env == "" {
+		if env == "" || m.IsEnvironment() {
 			continue
 		}
 		if !envName.MatchString(env) {

internal/control/catalog/shared.go

@@ -0,0 +1,110 @@
+package catalog
+
+import (
+	"time"
+)
+
+// A shared replica: one ComfyUI serving every model of an environment.
+//
+// Each ComfyUI model ran in a replica of its own, so on one card a picture
+// model and a 3D model took turns, each stopping the other and loading
+// ComfyUI and its weights from nothing. An environment of two or more models
+// gets a hidden catalog entry of its own, whose replica carries every
+// member's weights and custom nodes and answers for each of them; the
+// members get no replicas of their own. Placement, idleness, eviction and
+// adoption then work as for any model, on that entry.
+
+// RoleEnvironment is the role of an environment's hidden entry.
+const RoleEnvironment = "environment"
+
+// EnvModelPrefix names an environment's hidden entry: comfyui-env-<name>.
+const EnvModelPrefix = "comfyui-env-"
+
+// SharedComfy is what an environment's entry carries of its members.
+type SharedComfy struct {
+	Environment string         `json:"environment"`
+	Members     []SharedMember `json:"members"`
+}
+
+// SharedMember is one model a shared replica answers for.
+type SharedMember struct {
+	Name    string   `json:"name"`
+	Aliases []string `json:"aliases,omitempty"`
+	// Source is the member's weights, which the replica's node fetches.
+	Source string `json:"source"`
+	// HFRepo is the repository its node loads the weights by name from.
+	HFRepo string `json:"hf_repo,omitempty"`
+	// Workflow and Checkpoint are what a request for it is sent.
+	Workflow   string `json:"workflow"`
+	Checkpoint string `json:"checkpoint"`
+}
+
+// IsEnvironment reports whether a model is an environment's hidden entry.
+func (m Model) IsEnvironment() bool { return m.Role == RoleEnvironment }
+
+// SharedBy is the hidden entry whose replica serves a model, when it is in
+// an environment of two or more.
+func (c *Catalog) SharedBy(model string) (string, bool) {
+	m, ok := c.Resolve(model)
+	if !ok || m.IsEnvironment() || m.ComfyEnvironment() == "" {
+		return "", false
+	}
+	name := EnvModelPrefix + m.ComfyEnvironment()
+	if _, ok := c.models[name]; !ok {
+		return "", false
+	}
+	return name, true
+}
+
+// envModel is the hidden entry for an environment of two or more models: on
+// the first member's profiles, needing the most memory any of them does,
+// on demand for the longest idle timeout of theirs.
+func envModel(env Environment, members []Model) Model {
+	first := members[0]
+	profs := make([]Profile, len(first.Profiles))
+	copy(profs, first.Profiles)
+	mem := 0
+	for _, m := range members {
+		for _, p := range m.Profiles {
+			mem = max(mem, p.MinMemoryGB)
+		}
+	}
+	shared := &SharedComfy{Environment: env.Name}
+	var idle time.Duration
+	minN, maxN := 0, 1
+	class := ""
+	for _, m := range members {
+		workflow, ckpt := m.ComfyWorkflow()
+		hf := ""
+		if m.ComfyUI != nil {
+			hf = m.ComfyUI.HuggingFaceRepo
+		}
+		shared.Members = append(shared.Members, SharedMember{Name: m.Name, Aliases: m.Aliases,
+			Source: m.Source, HFRepo: hf, Workflow: workflow, Checkpoint: ckpt})
+		idle = max(idle, m.Scaling.IdleTimeout)
+		minN, maxN = max(minN, m.Scaling.Min), max(maxN, m.Scaling.Max)
+		if ClassRank(m.EffectiveClass()) > ClassRank(class) || class == "" {
+			class = m.EffectiveClass()
+		}
+	}
+	for i := range profs {
+		profs[i].MinMemoryGB = max(profs[i].MinMemoryGB, mem)
+		profs[i].ComfyUI = nil // the environment's nodes, below
+		profs[i].Source = ""
+	}
+	return Model{
+		Name:        EnvModelPrefix + env.Name,
+		Role:        RoleEnvironment,
+		Description: "the shared ComfyUI replica of environment " + env.Name,
+		Source:      first.Source,
+		ComfyUI:     &ComfyUI{Environment: env.Name, Nodes: env.Nodes, Output: OutputEnvironment},
+		Profiles:    profs,
+		Scaling:     Scaling{Min: minN, Max: maxN, IdleTimeout: idle},
+		Class:       class,
+		Shared:      shared,
+	}
+}
+
+// OutputEnvironment is an environment entry's output: it makes nothing of
+// its own, its members do.
+const OutputEnvironment = "environment"

internal/control/catalog/shared_test.go

@@ -0,0 +1,51 @@
+package catalog
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+	"time"
+)
+
+func TestAnEnvironmentOfTwoGetsOneHiddenEntryThatServesBoth(t *testing.T) {
+	// flux and a mesh model on one card took turns, each loading ComfyUI
+	// from nothing; in one environment one replica answers for both.
+	flux := comfyModel("/w/flux.safetensors", &ComfyUI{Environment: "art", Workflow: "flux-schnell"})
+	flux.Scaling = Scaling{Max: 1, IdleTimeout: 15 * time.Minute}
+	sf3d := inEnv("sf3d", "art", "/w/sf3d/", CustomNode{Repo: "https://example.com/sf3d", Commit: commitA})
+	sf3d.ComfyUI.HuggingFaceRepo = "stabilityai/stable-fast-3d"
+	sf3d.ComfyUI.Checkpoint = "model.safetensors"
+	sf3d.Scaling = Scaling{Max: 1, IdleTimeout: 10 * time.Minute}
+	cat, err := New([]Model{flux, sf3d, inEnv("alone", "solo", "/w/alone.ckpt")})
+	if err != nil {
+		t.Fatal(err)
+	}
+	host, ok := cat.SharedBy("sf3d")
+	if !ok || host != "comfyui-env-art" {
+		t.Fatalf("sf3d shared by %q %v", host, ok)
+	}
+	if _, ok := cat.SharedBy("alone"); ok {
+		t.Error("an environment of one got a hidden entry")
+	}
+	em, _ := cat.Resolve(host)
+	if !em.IsEnvironment() || em.MakesImages() || em.MakesMeshes() || em.Scaling.IdleTimeout != 15*time.Minute ||
+		len(em.Shared.Members) != 2 || em.Shared.Members[1].HFRepo != "stabilityai/stable-fast-3d" {
+		t.Errorf("entry %+v", em)
+	}
+	p, ok := em.EngineProfile(em.Profiles[0].Name)
+	if !ok {
+		t.Fatal("no engine profile")
+	}
+	var ms []map[string]string
+	if err := json.Unmarshal([]byte(p.Env[SharedModelsEnv]), &ms); err != nil || len(ms) != 2 || ms[1]["source"] != "/w/sf3d/" {
+		t.Errorf("members for the agent %q %v", p.Env[SharedModelsEnv], err)
+	}
+	if !strings.Contains(p.Env[NodesEnv], "sf3d") {
+		t.Errorf("nodes %q", p.Env[NodesEnv])
+	}
+	// A catalog built again from its own models is the same catalog.
+	again, err := New(cat.Models())
+	if err != nil || len(again.Models()) != len(cat.Models()) {
+		t.Errorf("rebuilt: %v, %d models vs %d", err, len(again.Models()), len(cat.Models()))
+	}
+}

internal/control/control.go

@@ -371,6 +371,35 @@ func (c *Controller) endpoints() []endpoint.Info {
 			Weight:   rep.Weight,
 			Priority: prof.Priority,
 		}
+		if m.Shared != nil {
+			// One replica answering for every member of its environment:
+			// offered under their names, each sent its own workflow and
+			// checkpoint. Its own hidden name nobody asks for.
+			info.Models = nil
+			info.Comfy = map[string]endpoint.ComfyRoute{}
+			ok := true
+			for _, sm := range m.Shared.Members {
+				w, err := c.workflow(sm.Workflow)
+				if err != nil {
+					c.log.Warn("a shared ComfyUI member's workflow could not be read; its replica is not offered",
+						"model", sm.Name, "replica", rep.ID, "err", err)
+					ok = false
+					break
+				}
+				route := endpoint.ComfyRoute{Workflow: w, Checkpoint: sm.Checkpoint}
+				for _, n := range append([]string{sm.Name}, sm.Aliases...) {
+					info.Models = append(info.Models, n)
+					info.Comfy[n] = route
+				}
+				if info.Workflow == "" {
+					info.Workflow, info.UpstreamModel = w, sm.Checkpoint
+				}
+			}
+			if ok {
+				out = append(out, info)
+			}
+			continue
+		}
 		if prof.Engine == catalog.EngineComfyUI {
 			workflow, ckpt := m.ComfyWorkflowFor(prof)
 			w, err := c.workflow(workflow)

internal/control/modelserve/modelserve.go

@@ -490,6 +490,20 @@ func (s *Server) run(j *Job) {
 		// third busy.
 		req.ContextLength, req.CtxPool = embedSlotTokens, embedSlotTokens*embedSlots
 	}
+	// On one card, the weights and the context's cache have to fit it: a
+	// 70B with a 65536-token pool on a 24 GB card loaded its weights and then
+	// failed to make its context, every time, as did an 8B with 131072.
+	if len(layout) > 0 && layout[0].Class != "" && !layout[0].CPU {
+		fitted, why, err := fitPoolToCard(req, header, set.Bytes, layout[0].Class, max(layout[0].GPUsPerNode, 1))
+		if err != nil {
+			s.fail(j, "check", err)
+			return
+		}
+		if why != "" {
+			s.step(j, "check", Running, why)
+			req = fitted
+		}
+	}
 	j.Request = req
 	memGB := int((header.MemoryBytes(set.Bytes, req.CtxPool) + 1<<30 - 1) >> 30)
 	basis := fmt.Sprintf("weights %.1f GB + KV cache %.1f GB for %d tokens + engine",
@@ -1071,3 +1085,30 @@ func (s *Server) embeds(name string) bool {
 	}
 	return false
 }
+
+// fitPoolToCard shrinks a context pool the cards cannot hold to one they
+// can, saying so, and refuses weights that do not fit at all. The slots stay
+// as many as asked while each keeps 2048 tokens, then fewer.
+func fitPoolToCard(req Request, h gguf.Header, weights int64, class string, gpus int) (Request, string, error) {
+	maxTok := h.PoolThatFits(weights, gguf.ClassMemGB(class), gpus)
+	switch {
+	case maxTok == 0 || req.CtxPool <= maxTok:
+		return req, "", nil
+	case maxTok < 0:
+		return req, "", fmt.Errorf("its weights (%.1f GB) do not fit %d card(s) of %s with room to run: "+
+			"choose a smaller quantisation, or more cards", gb(weights), gpus, class)
+	}
+	slots := max(1, req.CtxPool/max(req.ContextLength, 1))
+	for slots > 1 && maxTok/slots < 2048 {
+		slots--
+	}
+	per := (maxTok / slots) / 1024 * 1024
+	if per < 1024 {
+		return req, "", fmt.Errorf("its weights (%.1f GB) leave room on %s for a context of %d tokens only",
+			gb(weights), class, maxTok)
+	}
+	was := req.CtxPool
+	req.ContextLength, req.CtxPool = per, per*slots
+	return req, fmt.Sprintf("a %d-token pool does not fit %s beside %.1f GB of weights: %d slots of %d tokens instead",
+		was, class, gb(weights), slots, per), nil
+}

internal/control/modelserve/modelserve_test.go

@@ -5,6 +5,7 @@ import (
 	"context"
 	"encoding/binary"
 	"fmt"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/gguf"
 	"net/http"
 	"net/http/httptest"
 	"slices"
@@ -412,3 +413,18 @@ func TestAnEmbeddingModelIsServedForEmbedding(t *testing.T) {
 		t.Errorf("a chat model got %q", args)
 	}
 }
+
+func TestAPoolTheCardCannotHoldIsShrunkNotStarted(t *testing.T) {
+	h := gguf.Header{Architecture: "apertus", BlockCount: 32, HeadCount: 32, HeadCountKV: 8, EmbeddingLength: 4096}
+	req := Request{CtxPool: 131072, ContextLength: 32768}
+	got, why, err := fitPoolToCard(req, h, int64(8.5*(1<<30)), "6000-24g", 1)
+	if err != nil || why == "" || got.CtxPool > 105000 || got.CtxPool/got.ContextLength != 4 {
+		t.Errorf("fitted %+v %q %v", got, why, err)
+	}
+	if _, _, err := fitPoolToCard(req, h, int64(30*(1<<30)), "6000-24g", 1); err == nil {
+		t.Error("weights larger than the card were not refused")
+	}
+	if got, why, _ := fitPoolToCard(Request{CtxPool: 16384, ContextLength: 4096}, h, int64(8.5*(1<<30)), "6000-24g", 1); why != "" || got.CtxPool != 16384 {
+		t.Errorf("a pool that fits was changed: %+v %q", got, why)
+	}
+}

internal/control/reconciler/demand.go

@@ -55,7 +55,7 @@ func (r *Reconciler) IdleFor(id string) (time.Duration, bool) {
 // from "there is genuinely nothing serving this", which are different answers
 // to give a client.
 func (r *Reconciler) Want(model string) bool {
-	m, ok := r.cat.Load().Resolve(model)
+	m, ok := r.cat.Load().Resolve(r.hostOf(model))
 	if !ok || !m.Scaling.OnDemand() {
 		return false
 	}
@@ -127,6 +127,9 @@ func (r *Reconciler) baseFor(m catalog.Model) int {
 	if m.IsExternal() {
 		return 0 // served elsewhere: nothing to place, whatever its scaling says
 	}
+	if _, shared := r.cat.Load().SharedBy(m.Name); shared {
+		return 0 // its environment's replica serves it
+	}
 	if lo := r.minFor(m); !m.Scaling.OnDemand() || lo > 0 {
 		return lo
 	}
@@ -199,6 +202,17 @@ type Load struct {
 // were measured to hold, so scaling down does not land on the edge.
 func (r *Reconciler) RecordLoad(loads map[string]Load) {
 	now := r.now()
+	// A shared replica's load is its members' together.
+	for name, l := range loads {
+		if host := r.hostOf(name); host != name {
+			h := loads[host]
+			h.Replicas = max(h.Replicas, l.Replicas)
+			h.Running += l.Running
+			h.Waiting += l.Waiting
+			loads[host] = h
+			delete(loads, name)
+		}
+	}
 	// The profile each model's replicas run on, for its measurement; read
 	// before the loop, which holds the lock.
 	running := map[string]string{}
@@ -280,3 +294,12 @@ func (r *Reconciler) RecordLoad(loads map[string]Load) {
 // after a scale-down may be. Below one so that the load that was there a
 // minute ago does not put them straight back over.
 const scaleDownHeadroom = 0.7
+
+// hostOf is the model whose replicas serve name: its environment's hidden
+// entry when it shares one, else itself.
+func (r *Reconciler) hostOf(name string) string {
+	if host, ok := r.cat.Load().SharedBy(name); ok {
+		return host
+	}
+	return name
+}

internal/control/reconciler/evict.go

@@ -88,6 +88,8 @@ func (r *Reconciler) forgetWantIfGone(model, stopped string) {
 // longest idle first.
 func (r *Reconciler) evictable(wantedBy string) []evictionCandidate {
 	cat := r.cat.Load()
+	// A member never evicts the replica that serves it.
+	wantedBy = r.hostOf(wantedBy)
 	now := r.now()
 
 	wantRank := -1

internal/control/reconciler/health.go

@@ -130,7 +130,7 @@ func (r *Reconciler) ModelHealth() []ModelHealth {
 // Unservable says why a model cannot be served at all, so a request for it is
 // told now rather than asked to retry something that will never come up.
 func (r *Reconciler) Unservable(model string) (string, bool) {
-	m, ok := r.cat.Load().Resolve(model)
+	m, ok := r.cat.Load().Resolve(r.hostOf(model))
 	if !ok {
 		return "", false
 	}
@@ -176,7 +176,7 @@ func (r *Reconciler) Unservable(model string) (string, bool) {
 // on another model; eta is zero when when the holder lets go is not known.
 func (r *Reconciler) WaitFor(model string) (heldBy string, eta time.Duration, ok bool) {
 	cat := r.cat.Load()
-	m, found := cat.Resolve(model)
+	m, found := cat.Resolve(r.hostOf(model))
 	if !found || !m.Scaling.OnDemand() {
 		return "", 0, false
 	}

internal/control/reconciler/shared_test.go

@@ -0,0 +1,57 @@
+package reconciler
+
+import (
+	"testing"
+	"time"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/control/catalog"
+)
+
+func sharedCatalog(t *testing.T) *catalog.Catalog {
+	t.Helper()
+	mk := func(name, out string) catalog.Model {
+		return catalog.Model{Name: name, Source: "/w/" + name + ".safetensors",
+			ComfyUI:  &catalog.ComfyUI{Environment: "art", Workflow: "flux-schnell", Output: out},
+			Scaling:  catalog.Scaling{Max: 1, IdleTimeout: time.Hour},
+			Profiles: []catalog.Profile{{Name: "p", Engine: catalog.EngineComfyUI, GPUClass: "h100-80g", TP: 1}}}
+	}
+	cat, err := catalog.New([]catalog.Model{mk("flux", catalog.OutputPictures), mk("sf3d", catalog.OutputMeshes)})
+	if err != nil {
+		t.Fatal(err)
+	}
+	return cat
+}
+
+func TestAMemberAskedForStartsItsEnvironmentsOneReplica(t *testing.T) {
+	// A picture model and a 3D model on one card took turns, each from
+	// nothing: in one environment, asking for either starts one replica
+	// that serves both, and neither gets one of its own.
+	h := newHarness(t, testModel(1), gpuNode("1", 1))
+	h.r.cat.Store(sharedCatalog(t))
+	if !h.r.Want("sf3d") {
+		t.Fatal("a member was not wanted")
+	}
+	h.r.Reconcile(t.Context())
+	h.markReady()
+	h.r.Want("flux")
+	h.r.Reconcile(t.Context())
+	var models []string
+	for _, rep := range h.r.Replicas() {
+		if rep.State.Counts() {
+			models = append(models, rep.Model)
+		}
+	}
+	if len(models) != 1 || models[0] != "comfyui-env-art" {
+		t.Fatalf("replicas %v", models)
+	}
+	// The member's sibling never evicts the replica serving it.
+	reps := h.r.Replicas()
+	h.r.RecordUsage(map[string]time.Time{reps[0].ID: h.clock.now()})
+	h.advance(DefaultConfig().MinLifetime + DefaultConfig().EvictionGrace + time.Minute)
+	if got := h.r.evictable("flux"); len(got) != 0 {
+		t.Errorf("a member was offered its own environment's replica: %+v", got)
+	}
+	if _, _, ok := h.r.WaitFor("flux"); ok {
+		t.Error("a member served by a replica that is up was said to be waiting")
+	}
+}

internal/gguf/gguf.go

@@ -378,3 +378,40 @@ func Check(ctx context.Context, url, llamaVersion string) Verdict {
 	v.Header = h
 	return v
 }
+
+// CardUsable is the share of a card's memory a model may take: the rest is
+// CUDA's own and the engine's compute buffers.
+const CardUsable = 0.9
+
+// PoolThatFits is the largest context pool, in tokens, whose KV cache fits
+// on gpus cards of cardGB beside weights bytes of weights and a gigabyte for
+// the engine: -1 when the weights alone do not, 0 when the header or the
+// card does not say enough to tell.
+func (h Header) PoolThatFits(weights int64, cardGB, gpus int) int {
+	perTok := h.KVBytesPerToken()
+	if cardGB <= 0 || perTok <= 0 || weights <= 0 {
+		return 0
+	}
+	budget := float64(cardGB*max(gpus, 1))*CardUsable*(1<<30) - float64(weights) - (1 << 30)
+	if budget <= 0 {
+		return -1
+	}
+	return int(budget / float64(perTok))
+}
+
+// ClassMemGB is a card's memory as its class names it: 6000-24g is 24, 0
+// for a class that does not say.
+func ClassMemGB(class string) int {
+	i := strings.LastIndex(class, "-")
+	if i < 0 || !strings.HasSuffix(class, "g") {
+		return 0
+	}
+	n := 0
+	for _, c := range strings.TrimSuffix(class[i+1:], "g") {
+		if c < '0' || c > '9' {
+			return 0
+		}
+		n = n*10 + int(c-'0')
+	}
+	return n
+}

internal/gguf/gguf_test.go

@@ -188,3 +188,25 @@ func TestAnEmbeddingModelIsKnownByItsPooling(t *testing.T) {
 		t.Errorf("no pooling: %+v %v", h, err)
 	}
 }
+
+func TestAPoolIsSizedToTheCard(t *testing.T) {
+	// Apertus-8B's shape: 32 layers, 8 KV heads of 128, 128 KiB a token.
+	h, err := Read(bytes.NewReader(shapedHeader("apertus", 32, 32, 8, 4096, true)))
+	if err != nil {
+		t.Fatal(err)
+	}
+	weights := int64(8.5 * (1 << 30))
+	most := h.PoolThatFits(weights, ClassMemGB("6000-24g"), 1)
+	if most < 90000 || most > 105000 {
+		t.Errorf("an 8B on 24 GB holds %d tokens of pool", most)
+	}
+	if most >= 131072 {
+		t.Error("the 131072 pool that failed on the card would pass")
+	}
+	if got := h.PoolThatFits(int64(30*(1<<30)), 24, 1); got != -1 {
+		t.Errorf("30 GB of weights on 24 GB: %d", got)
+	}
+	if ClassMemGB("t4-15g") != 15 || ClassMemGB("cpu") != 0 || ClassMemGB("v100-32g") != 32 {
+		t.Error("class memory misread")
+	}
+}

internal/notify/notify.go

@@ -258,7 +258,11 @@ type Result struct {
 
 // redact keeps a webhook's host and drops its path, where Slack and others
 // put the secret.
-func redact(u string) string {
+func redact(u string) string { return Redact(u) }
+
+// Redact is a webhook's address as it may be shown: its host, not the path
+// that is its secret.
+func Redact(u string) string {
 	if i := strings.Index(u, "://"); i >= 0 {
 		if j := strings.IndexByte(u[i+3:], '/'); j >= 0 {
 			return u[:i+3+j] + "/…"

internal/router/canary.go

@@ -111,7 +111,7 @@ func (r *Router) isCanaryTarget(name string) bool {
 
 // hiddenModel is a model clients cannot name: a judge, or a canary.
 func (r *Router) hiddenModel(name string) bool {
-	return r.isGuardModel(name) || r.isCanaryTarget(name)
+	return r.isGuardModel(name) || r.isCanaryTarget(name) || r.modelMap()[name].IsEnvironment()
 }
 
 // setVariantHeader tells the client which side served it.

internal/router/endpoint/endpoint.go

@@ -51,6 +51,25 @@ type Info struct {
 	// Workflow is the ComfyUI workflow sent for an endpoint whose engine is
 	// EngineComfyUI: a built-in's name or the workflow's JSON.
 	Workflow string
+	// Comfy is, for a shared ComfyUI replica that answers for several
+	// models, each one's workflow and checkpoint, by every name it goes by.
+	// Nil for a replica of one model, which Workflow and UpstreamModel say.
+	Comfy map[string]ComfyRoute
+}
+
+// ComfyRoute is what a request for one model of a shared ComfyUI replica is
+// sent.
+type ComfyRoute struct {
+	Workflow   string
+	Checkpoint string
+}
+
+// ComfyFor is the workflow and checkpoint a request for model is sent here.
+func (i Info) ComfyFor(model string) (workflow, checkpoint string) {
+	if r, ok := i.Comfy[model]; ok {
+		return r.Workflow, r.Checkpoint
+	}
+	return i.Workflow, i.UpstreamModel
 }
 
 // EngineExternal marks an endpoint that is a provider outside the cluster.

internal/router/files.go

@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"io"
 	"net/http"
+	"slices"
 	"time"
 
 	"gitea.mynodes.xyz/ilker/llmgw/internal/auth"
@@ -20,8 +21,16 @@ import (
 const (
 	purposeBatch       = "batch"
 	purposeBatchOutput = "batch_output"
+	// purposeVision and purposeUserData are files a request works from — a
+	// picture to make a mesh of or to look at, a workflow's input — named
+	// by id rather than sent as base64 in every call. OpenAI's names.
+	purposeVision   = "vision"
+	purposeUserData = "user_data"
 )
 
+// uploadPurposes are what a client may upload a file for.
+var uploadPurposes = []string{purposeBatch, purposeVision, purposeUserData}
+
 func newID(prefix string) string {
 	var b [12]byte
 	_, _ = rand.Read(b[:])
@@ -99,9 +108,11 @@ func (r *Router) handleUploadFile(w http.ResponseWriter, req *http.Request) {
 			fmt.Sprintf("a multipart form with purpose and file, at most %d MB: %v", maxBytes>>20, err))
 		return
 	}
-	if purpose := req.FormValue("purpose"); purpose != purposeBatch {
+	purpose := req.FormValue("purpose")
+	if !slices.Contains(uploadPurposes, purpose) {
 		writeError(w, http.StatusBadRequest, "invalid_request_error",
-			fmt.Sprintf("purpose %q: this gateway keeps files for batches only (purpose batch)", purpose))
+			fmt.Sprintf("purpose %q: this gateway keeps files for batches (batch) and for requests to work from "+
+				"(vision, user_data)", purpose))
 		return
 	}
 	file, header, err := req.FormFile("file")
@@ -120,8 +131,13 @@ func (r *Router) handleUploadFile(w http.ResponseWriter, req *http.Request) {
 			fmt.Sprintf("the file is larger than %d MB", maxBytes>>20))
 		return
 	}
-	f := store.File{ID: newID("file-"), Owner: ownerOf(p), Purpose: purposeBatch,
+	f := store.File{ID: newID("file-"), Owner: ownerOf(p), Purpose: purpose,
 		Filename: header.Filename, Bytes: int64(len(content)), CreatedAt: time.Now()}
+	// A file to work from is kept as long as what is made from it.
+	if purpose != purposeBatch {
+		_, _, _, days := r.batchCfg()
+		f.ExpiresAt = f.CreatedAt.AddDate(0, 0, days)
+	}
 	if err := r.store.PutFile(req.Context(), f, content); err != nil {
 		writeError(w, http.StatusInternalServerError, "server_error", err.Error())
 		return

internal/router/images.go

@@ -182,7 +182,8 @@ func (r *Router) generateImages(ctx context.Context, w http.ResponseWriter, req
 		Model string `json:"model"`
 	}
 	_ = json.Unmarshal(body, &upstream)
-	tpl, err := comfy.Load(ep.Workflow)
+	wf, _ := ep.ComfyFor(req.Model)
+	tpl, err := comfy.Load(wf)
 	if err != nil {
 		return proxy.Outcome{}, fmt.Errorf("%s: %w", ep.ID, err)
 	}

internal/router/mcp_media.go

@@ -348,6 +348,45 @@ func (r *Router) selfJSON(c mediaCall, method, path string, body, into any) (htt
 	return h, json.Unmarshal(b, into)
 }
 
+// selfJSONWaiting is selfJSON for a model that may be cold: an answer that
+// says to come back — a model loading, another holding the GPU, a full
+// queue, all with Retry-After — is waited out for as long as the call may
+// wait, saying so meanwhile. One that does not say when, a spent budget, is
+// not: waiting would not help.
+func (r *Router) selfJSONWaiting(c mediaCall, model, path string, body, into any) (http.Header, error) {
+	deadline := time.Now().Add(c.wait(nil))
+	for {
+		status, h, b, err := r.self(c, http.MethodPost, path, body)
+		if err != nil {
+			return nil, err
+		}
+		retry, _ := strconv.Atoi(h.Get("Retry-After"))
+		if status < 300 {
+			return h, json.Unmarshal(b, into)
+		}
+		if retry <= 0 || (status != http.StatusTooManyRequests && status < 500) || !time.Now().Before(deadline) {
+			msg := fmt.Sprintf("POST %s: %d %s", path, status, errorMessage(b))
+			if retry > 0 {
+				msg += "; still not ready after waiting — call again, or ask gateway_status which model holds the GPU"
+			}
+			return h, errors.New(msg)
+		}
+		say := model + " is loading"
+		if held := h.Get(waitingForHeader); held != "" {
+			say = model + " is waiting for " + held + " to give up the GPU"
+			if eta := h.Get(etaHeader); eta != "" {
+				say += ", about " + eta + " s"
+			}
+		}
+		c.progress(5, say)
+		select {
+		case <-c.req.Context().Done():
+			return nil, c.req.Context().Err()
+		case <-time.After(min(time.Duration(retry)*time.Second, 10*time.Second, time.Until(deadline)+time.Second)):
+		}
+	}
+}
+
 // fileURL is where a file is fetched from, with the caller's key, as the
 // caller reached the gateway.
 func fileURL(req *http.Request, id string) string {
@@ -518,7 +557,7 @@ func (r *Router) makeImage(c mediaCall, prompt, style, negative, model string, s
 			B64 string `json:"b64_json"`
 		} `json:"data"`
 	}
-	h, err := r.selfJSON(c, http.MethodPost, imagesPath, body, &out)
+	h, err := r.selfJSONWaiting(c, model, imagesPath, body, &out)
 	if err != nil {
 		return nil, nil, err
 	}
@@ -614,6 +653,10 @@ func (r *Router) toolGenerateMesh(c mediaCall, raw json.RawMessage) (map[string]
 			return nil, fmt.Errorf("%s's workflow takes no picture", a.Model)
 		}
 		inputs := map[string]any{image: dataURL(pic)}
+		// A file of the caller's goes by its id, which the API takes.
+		if id := strings.TrimSpace(a.Image); strings.HasPrefix(id, "file-") {
+			inputs[image] = id
+		}
 		for k, v := range a.Inputs {
 			inputs[k] = v
 		}
@@ -827,7 +870,7 @@ func (r *Router) toolReviewImages(c mediaCall, raw json.RawMessage) (map[string]
 			} `json:"message"`
 		} `json:"choices"`
 	}
-	if _, err := r.selfJSON(c, http.MethodPost, "/v1/chat/completions", body, &out); err != nil {
+	if _, err := r.selfJSONWaiting(c, a.Model, "/v1/chat/completions", body, &out); err != nil {
 		return nil, err
 	}
 	if len(out.Choices) == 0 {

internal/router/mcp_media_test.go

@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"image"
 	"image/png"
+	"mime/multipart"
 	"net/http"
 	"net/http/httptest"
 	"strings"
@@ -264,3 +265,107 @@ func TestPicturesAreLookedAtByAVisionModelShrunkAndToASchema(t *testing.T) {
 		t.Errorf("the picture sent is %v (%v), not shrunk to %d", im.Bounds(), err, lookEdge)
 	}
 }
+
+// noisePNG is a picture that does not compress: n by n, so its base64 is
+// past any small limit.
+func noisePNG(t *testing.T, n int) []byte {
+	t.Helper()
+	im := image.NewRGBA(image.Rect(0, 0, n, n))
+	seed := uint32(1)
+	for i := range im.Pix {
+		seed = seed*1664525 + 1013904223
+		im.Pix[i] = byte(seed >> 24)
+	}
+	var b bytes.Buffer
+	_ = png.Encode(&b, im)
+	return b.Bytes()
+}
+
+func TestAPictureOfRealSizeGoesThroughMCPAndByFileID(t *testing.T) {
+	c := newMeshComfy(t)
+	rt, secret := meshRouter(t, c)
+	h := rt.Handler()
+	pic := noisePNG(t, 400) // ~640 KB, as a 1024 px concept is
+	res, _ := mcp(t, h, secret, "tools/call", map[string]any{"name": "generate_mesh", "arguments": map[string]any{
+		"model": "trellis2", "image": "data:image/png;base64," + base64.StdEncoding.EncodeToString(pic)}}, false)
+	if info := resultInfo(t, res); info["status"] != "completed" {
+		t.Fatalf("a mesh from a picture of real size: %v", info)
+	}
+
+	// Uploaded once as a picture to work from, then named by id.
+	var form bytes.Buffer
+	mw := multipart.NewWriter(&form)
+	_ = mw.WriteField("purpose", "vision")
+	fw, _ := mw.CreateFormFile("file", "rock.png")
+	_, _ = fw.Write(pic)
+	_ = mw.Close()
+	req := httptest.NewRequest(http.MethodPost, "/v1/files", &form)
+	req.Header.Set("Content-Type", mw.FormDataContentType())
+	req.Header.Set("Authorization", "Bearer "+secret)
+	rec := httptest.NewRecorder()
+	h.ServeHTTP(rec, req)
+	var f map[string]any
+	if rec.Code != http.StatusOK || json.Unmarshal(rec.Body.Bytes(), &f) != nil || f["purpose"] != "vision" ||
+		f["expires_at"] == nil {
+		t.Fatalf("a picture uploaded to work from: %d %s", rec.Code, rec.Body.String())
+	}
+	res, _ = mcp(t, h, secret, "tools/call", map[string]any{"name": "generate_mesh", "arguments": map[string]any{
+		"model": "trellis2", "image": f["id"]}}, false)
+	if info := resultInfo(t, res); info["status"] != "completed" {
+		t.Fatalf("a mesh from an uploaded picture: %v", info)
+	}
+}
+
+func TestACallTooLargeIsSaidSo(t *testing.T) {
+	h := newTestRouter(t, baseConfig()).Handler()
+	big := `{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"x","arguments":{"image":"` +
+		strings.Repeat("A", mcpBody) + `"}}}`
+	req := httptest.NewRequest(http.MethodPost, "/mcp", strings.NewReader(big))
+	rec := httptest.NewRecorder()
+	h.ServeHTTP(rec, req)
+	if rec.Code != http.StatusRequestEntityTooLarge || !strings.Contains(rec.Body.String(), "larger than") {
+		t.Errorf("%d %s", rec.Code, rec.Body.String()[:min(200, rec.Body.Len())])
+	}
+}
+
+func TestACallToAColdModelWaitsItOutSayingSo(t *testing.T) {
+	rt := newTestRouter(t, baseConfig())
+	_ = rt.Handler()
+	calls := 0
+	var fake http.Handler = http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		calls++
+		if calls < 3 {
+			w.Header().Set("Retry-After", "1")
+			w.Header().Set(waitingForHeader, "sf3d")
+			w.Header().Set(etaHeader, "40")
+			writeError(w, http.StatusServiceUnavailable, "server_error", "model loading: waiting")
+			return
+		}
+		w.Header().Set("x-llmgw-seed", "4")
+		writeJSON(w, http.StatusOK, map[string]any{"ok": true})
+	})
+	rt.selfHandler.Store(&fake)
+	var said []string
+	c := mediaCall{req: httptest.NewRequest(http.MethodPost, "/mcp", nil), streamed: true,
+		progress: func(_ float64, m string) { said = append(said, m) }}
+	var out map[string]any
+	h, err := rt.selfJSONWaiting(c, "flux-schnell", imagesPath, map[string]any{}, &out)
+	if err != nil || out["ok"] != true || h.Get("x-llmgw-seed") != "4" || calls != 3 {
+		t.Fatalf("%v %v after %d calls", err, out, calls)
+	}
+	if len(said) != 2 || !strings.Contains(said[0], "waiting for sf3d to give up the GPU, about 40 s") {
+		t.Errorf("said %q", said)
+	}
+
+	// Refused for good — no Retry-After — is said at once.
+	calls = 0
+	fake = http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		calls++
+		writeError(w, http.StatusPaymentRequired, "insufficient_quota", "the budget is spent")
+	})
+	rt.selfHandler.Store(&fake)
+	if _, err := rt.selfJSONWaiting(c, "flux-schnell", imagesPath, map[string]any{}, &out); err == nil ||
+		!strings.Contains(err.Error(), "budget") || calls != 1 {
+		t.Errorf("%v after %d calls", err, calls)
+	}
+}

internal/router/memory.go

@@ -3,6 +3,7 @@ package router
 import (
 	"encoding/json"
 	"errors"
+	"fmt"
 	"io"
 	"net/http"
 	"slices"
@@ -121,8 +122,7 @@ func (r *Router) handlePutNote(w http.ResponseWriter, req *http.Request) {
 		Text string   `json:"text"`
 		Tags []string `json:"tags"`
 	}
-	// Pictures to look at come in the call, as base64.
-	raw, err := io.ReadAll(http.MaxBytesReader(w, req.Body, mcpBody))
+	raw, err := io.ReadAll(http.MaxBytesReader(w, req.Body, maxMemoryBody))
 	if err != nil || json.Unmarshal(raw, &body) != nil {
 		writeError(w, http.StatusBadRequest, "invalid_request_error",
 			`body must be JSON like {"text": "...", "tags": ["..."]}`)
@@ -257,7 +257,17 @@ var mcpTools = []map[string]any{
 
 // handleMCP serves POST /mcp.
 func (r *Router) handleMCP(w http.ResponseWriter, req *http.Request) {
-	raw, err := io.ReadAll(http.MaxBytesReader(w, req.Body, maxMemoryBody))
+	// Pictures to make a mesh from or to look at come in the call, as
+	// base64.
+	raw, err := io.ReadAll(http.MaxBytesReader(w, req.Body, mcpBody))
+	if mbe := (*http.MaxBytesError)(nil); errors.As(err, &mbe) {
+		writeJSON(w, http.StatusRequestEntityTooLarge, rpcResponse{
+			JSONRPC: "2.0", ID: json.RawMessage("null"),
+			Error: &rpcError{Code: -32600, Message: fmt.Sprintf("the call is larger than %d MB: send pictures smaller, "+
+				"or fewer at once", mcpBody>>20)},
+		})
+		return
+	}
 	var rpc rpcRequest
 	if err != nil || json.Unmarshal(raw, &rpc) != nil || rpc.Method == "" {
 		writeJSON(w, http.StatusBadRequest, rpcResponse{

internal/router/meshes.go

@@ -101,7 +101,8 @@ func (r *Router) generateMeshes(ctx context.Context, w http.ResponseWriter, req
 	if err != nil {
 		return proxy.Outcome{}, err
 	}
-	tpl, err := comfy.Load(ep.Workflow)
+	wf, _ := ep.ComfyFor(req.Model)
+	tpl, err := comfy.Load(wf)
 	if err != nil {
 		return proxy.Outcome{}, fmt.Errorf("%s: %w", ep.ID, err)
 	}

internal/router/router.go

@@ -1138,8 +1138,10 @@ func (r *Router) dispatchOne(ctx context.Context, w http.ResponseWriter, req *pa
 			trace.WithAttributes(attribute.String("llmgw.endpoint", dec.Endpoint.ID),
 				attribute.String("url.full", dec.Endpoint.URL+req.Path)))
 		body := req.UpstreamBody
-		if dec.Endpoint.UpstreamModel != "" && req.ContentType == "" {
-			body = withModel(body, dec.Endpoint.UpstreamModel)
+		// A shared ComfyUI replica is told which of its models' checkpoint;
+		// any other endpoint its own upstream name.
+		if _, up := dec.Endpoint.ComfyFor(model); up != "" && req.ContentType == "" {
+			body = withModel(body, up)
 		}
 		// llama-server is asked for the model with the adapter applied to
 		// this request; vLLM serves each adapter as a model of its own name,

internal/router/workflows.go

@@ -297,7 +297,8 @@ func (r *Router) generateRun(ctx context.Context, w http.ResponseWriter, req *pa
 	if err != nil {
 		return proxy.Outcome{}, err
 	}
-	tpl, err := comfy.Load(ep.Workflow)
+	wf, _ := ep.ComfyFor(req.Model)
+	tpl, err := comfy.Load(wf)
 	if err != nil {
 		return proxy.Outcome{}, fmt.Errorf("%s: %w", ep.ID, err)
 	}

internal/web/admin_mirror.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"encoding/json"
 	"errors"
+	"fmt"
 	"html/template"
 	"net/http"
 	"slices"
@@ -342,7 +343,14 @@ func (p *AccountPages) RegisterAdminMirror(mux *http.ServeMux, d MirrorDeps) {
 				// between the listing and the click, or a mirror that has gone
 				// down since, both end as a replica failing on a node minutes
 				// later with an error naming the node. Asking costs one HEAD.
-				if _, err := sourcecheck.Check(ctx, model.Source); err != nil {
+				size, err := sourcecheck.Check(ctx, model.Source)
+				if err != nil {
+					return "", err
+				}
+				// The weights and the context's cache have to fit the card:
+				// an 8B asked for a 131072-token pool on 24 GB loaded its
+				// weights and failed to make its context, again and again.
+				if err := p.fitsCard(ctx, l, model, size); err != nil {
 					return "", err
 				}
 				if err := d.Catalog.Add(ctx, model, admin.Email); err != nil {
@@ -365,6 +373,14 @@ func (p *AccountPages) RegisterAdminMirror(mux *http.ServeMux, d MirrorDeps) {
 	mux.HandleFunc("POST /admin/mirror/{model}/{version}/remove", p.mirrorAction(m,
 		func(ctx context.Context, _ store.User, r *http.Request) (string, error) {
 			model, ver := r.PathValue("model"), r.PathValue("version")
+			// Weights a catalog model still loads are not taken from under
+			// it: removed from the mirror alone, the model stayed in the
+			// catalog and its replicas went on loading the copy on the node.
+			if users := p.mirrorUsers(ctx, model, ver); len(users) > 0 {
+				return "", fmt.Errorf("%s/%s is the weights of %s in the catalog: remove %s first "+
+					"(the models page, or llmgw ctl model remove %s), then the weights",
+					model, ver, strings.Join(users, ", "), strings.Join(users, ", "), users[0])
+			}
 			removed, err := m.Remove(ctx, model, ver)
 			if err != nil {
 				return "", err
@@ -937,3 +953,52 @@ func sharedPrefix(ids []string) string {
 	}
 	return ""
 }
+
+// mirrorUsers are the catalog models whose weights are the mirror's
+// model/version.
+func (p *AccountPages) mirrorUsers(ctx context.Context, model, ver string) []string {
+	if p.catalog == nil {
+		return nil
+	}
+	srcs, err := p.catalog.Sources(ctx)
+	if err != nil {
+		return nil
+	}
+	part := "/" + model + "/" + ver + "/"
+	var out []string
+	for src, ms := range srcs {
+		if !strings.Contains(src, part) {
+			continue
+		}
+		for _, m := range ms {
+			if !slices.Contains(out, m.Name) {
+				out = append(out, m.Name)
+			}
+		}
+	}
+	slices.Sort(out)
+	return out
+}
+
+// fitsCard refuses a model whose weights and context pool do not fit the
+// cards its profile names, saying the largest pool that would.
+func (p *AccountPages) fitsCard(ctx context.Context, l *Lang, m catalog.Model, size int64) error {
+	if len(m.Profiles) == 0 || m.Profiles[0].GPUClass == "" || size <= 0 {
+		return nil
+	}
+	pr := m.Profiles[0]
+	v := p.checkSource(ctx, m.Source)
+	cards := max(pr.GPUsPerNode, pr.TP, 1)
+	most := v.Header.PoolThatFits(size, gguf.ClassMemGB(pr.GPUClass), cards)
+	pool := argInt(pr.Args, "--ctx-size")
+	if pool <= 0 {
+		pool = defaultCtxPool
+	}
+	switch {
+	case most < 0:
+		return errors.New(l.T("serve.err.weights_too_big", l.Num(float64(size)/(1<<30), 1), pr.GPUClass))
+	case most > 0 && pool > most:
+		return errors.New(l.T("serve.err.pool_too_big", l.Int(int64(pool)), pr.GPUClass, l.Int(int64(most/1024*1024))))
+	}
+	return nil
+}

internal/web/admin_mirror_test.go

@@ -766,3 +766,28 @@ func TestAJudgesWeightsAreServedAsTheJudgeThePoliciesAsk(t *testing.T) {
 		t.Errorf("a model clients use became a judge: %+v", plain)
 	}
 }
+
+func TestWeightsAModelStillLoadsAreNotRemoved(t *testing.T) {
+	// Removed from the mirror alone, a model stayed in the catalog and its
+	// replicas went on loading the copy on the node, with nothing said.
+	m := &fakeMirror{}
+	cat := &fakeCatalog{sources: map[string][]catalog.Model{
+		"http://10.0.0.146:8081/apertus-70b/aff98136/Apertus-70B-UD-IQ1_M.gguf": {{Name: "apertus-70b-instruct-2509"}},
+	}}
+	a := fakeAccounts{user: store.User{ID: "u1", Email: "a@example.com", IsAdmin: true}}
+	p := NewAccountPages(a, nil, nil, "v0", slog.New(slog.NewTextHandler(io.Discard, nil)))
+	mux := http.NewServeMux()
+	p.RegisterAdminMirror(mux, MirrorDeps{Mirror: m, Catalog: cat})
+	body := mirrorPost(t, mux, "/admin/mirror/apertus-70b/aff98136/remove", url.Values{"csrf": {"tok"}}).Body.String()
+	if m.removed != [2]string{} {
+		t.Errorf("weights in use were removed: %v", m.removed)
+	}
+	if !strings.Contains(body, "apertus-70b-instruct-2509") || !strings.Contains(body, "remove") {
+		t.Errorf("the page does not say which model uses them: %s", body)
+	}
+	// Weights nothing loads go as before.
+	mirrorPost(t, mux, "/admin/mirror/other/v1/remove", url.Values{"csrf": {"tok"}})
+	if m.removed != [2]string{"other", "v1"} {
+		t.Errorf("unused weights were not removed: %v", m.removed)
+	}
+}

internal/web/locales/en/serve.json

@@ -73,6 +73,8 @@
   "serve.err.engine": "which engine should run it?",
   "serve.err.needs_gpu": "%s runs on GPUs; choose a GPU class to place it on",
   "serve.err.placement": "%q is not a placement",
+  "serve.err.weights_too_big": "Its weights (%s GB) do not fit a %s card with room to run: choose a smaller quantisation, or more cards.",
+  "serve.err.pool_too_big": "A context pool of %s tokens does not fit a %s card beside these weights; at most %s would.",
   "serve.err.cpu_with_free_gpus.one": "this fleet has %d GPU free and %s was set to run on the CPU; choose a GPU placement, or tick \"run on the CPU anyway\"",
   "serve.err.cpu_with_free_gpus.other": "this fleet has %d GPUs free and %s was set to run on the CPU; choose a GPU placement, or tick \"run on the CPU anyway\"",
   "serve.msg.added": "%s added to the catalog, %s: %s; it is placed on the next pass",

internal/web/locales/tr/serve.json

@@ -72,6 +72,8 @@
   "serve.err.engine": "hangi motor çalıştıracak?",
   "serve.err.needs_gpu": "%s GPU üzerinde çalışır; yerleştirmek için bir GPU sınıfı seçin",
   "serve.err.placement": "%q geçerli bir yerleşim değil",
+  "serve.err.weights_too_big": "Ağırlıkları (%s GB) bir %s kartına çalışacak payla sığmıyor: daha küçük bir quantization ya da daha fazla kart seçin.",
+  "serve.err.pool_too_big": "%s token'lık bağlam havuzu bu ağırlıklarla bir %s kartına sığmıyor; en fazla %s sığar.",
   "serve.err.cpu_with_free_gpus.other": "bu kümede %d GPU boşta, %s ise CPU üzerinde çalışacak biçimde ayarlandı; bir GPU yerleşimi seçin ya da \"yine de CPU üzerinde çalıştır\" kutusunu işaretleyin",
   "serve.msg.added": "%s kataloğa eklendi, %s: %s; bir sonraki geçişte yerleştirilecek",
   "serve.replicas_n.other": "%d replika",