A model waiting for an on-demand one to step aside is waiting, not unplaceable

A 3D job asked for straight after a picture, on a cluster with one card,
failed at once with "cannot be placed": flux held the card, too freshly
started to evict. The cards coming back when an on-demand model goes idle
makes it a wait, so the request is told to retry and the job waits.

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

internal/control/reconciler/health.go

@@ -134,21 +134,39 @@ func (r *Reconciler) Unservable(model string) (string, bool) {
 	if !ok {
 		return "", false
 	}
+	cat := r.cat.Load()
 	r.mu.Lock()
-	defer r.mu.Unlock()
 	if p, ok := r.parked[m.Name]; ok {
+		r.mu.Unlock()
 		return fmt.Sprintf("%s failed to start %d times in a row and is not being retried: %s",
 			m.Name, p.Failures, p.Reason), true
 	}
+	var asleepLater []victim
 	for _, rep := range r.replicas {
-		if rep.Model == m.Name && rep.State.Counts() {
+		if !rep.State.Counts() {
+			continue
+		}
+		if rep.Model == m.Name {
+			r.mu.Unlock()
 			return "", false // something is on its way
 		}
+		if other, ok := cat.Resolve(rep.Model); ok && other.Scaling.OnDemand() && len(nodesOf(rep)) == 1 {
+			asleepLater = append(asleepLater, victim{ID: rep.ID, Model: rep.Model})
+		}
 	}
-	if res, ok := r.lastExplain[m.Name]; ok && res.Placement == nil {
-		return m.Name + " cannot be placed on any node: " + firstLine(res.Explain()), true
+	res, ok := r.lastExplain[m.Name]
+	r.mu.Unlock()
+	if !ok || res.Placement != nil {
+		return "", false
 	}
-	return "", false
+	// The cards are held by models that let go of them when idle, which
+	// makeRoomFor hastens once they have been up a while. That is a wait, not
+	// a never: a 3D model asked for straight after a picture, on a cluster
+	// with one card, waits for the picture model to step aside.
+	if m.Scaling.OnDemand() && len(asleepLater) > 0 && r.fitsWithout(m, nil, asleepLater) {
+		return "", false
+	}
+	return m.Name + " cannot be placed on any node: " + firstLine(res.Explain()), true
 }
 
 func firstLine(s string) string {

internal/control/reconciler/reconciler_test.go

@@ -1708,6 +1708,45 @@ func TestAFreshlyStartedReplicaIsNotEvicted(t *testing.T) {
 	}
 }
 
+func TestAModelWaitingForAnotherOnDemandOneToStepAsideIsNotUnservable(t *testing.T) {
+	// A 3D model asked for straight after a picture, on a cluster with one
+	// card: the picture model holds it, too freshly started to evict. That
+	// is a wait — told "cannot be placed", a job failed at once.
+	h := newHarness(t, testModel(1), gpuNode("1", 4))
+	h.r.cat.Store(twoOnDemandModels(t))
+	h.r.Want("llama")
+	h.r.Reconcile(t.Context())
+	h.markReady()
+	reps := h.r.Replicas()
+	if len(reps) != 1 {
+		t.Fatalf("setup: %d replicas", len(reps))
+	}
+	h.r.RecordUsage(map[string]time.Time{reps[0].ID: h.clock.now()})
+
+	h.r.Want("second")
+	h.r.Reconcile(t.Context())
+	if res, ok := h.r.lastExplain["second"]; !ok || res.Placement != nil {
+		t.Fatalf("setup: second was not left unplaced: %+v", res)
+	}
+	if why, no := h.r.Unservable("second"); no {
+		t.Errorf("a model waiting for an on-demand one to go idle was called unservable: %q", why)
+	}
+
+	// Held by a model that stays up, the card is not coming back.
+	always := testModel(1)
+	always.Name = "llama"
+	always.Scaling = catalog.Scaling{Min: 1, Max: 1}
+	second := mustResolve(t, h, "second")
+	built, err := catalog.New([]catalog.Model{always, second})
+	if err != nil {
+		t.Fatal(err)
+	}
+	h.r.cat.Store(built)
+	if _, no := h.r.Unservable("second"); !no {
+		t.Error("a model behind an always-on one was not called unservable")
+	}
+}
+
 func TestAStartedShardIsToldWhichProfileVersionItIs(t *testing.T) {
 	// The fingerprint has to leave the control plane, or it cannot come back
 	// after a restart — and a restart is exactly when the configuration is