A model that does not fit its card is not added to fail; weights in use are not removed

Apertus 70B with a 65536-token pool, and then 8B with 131072, on a 24 GB
card loaded their weights and failed to make their context on every start,
the 8B taking the card from the 70B each time only to fail. ctl model serve
now shrinks a pool the card cannot hold beside the weights (KV cache from
the GGUF's shape, 90% of the card, a gigabyte for the engine), 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.

The mirror page removed the 70B's weights while the catalog still had the
model, whose replicas went on loading the node's copy with nothing said.
Removing weights a catalog model loads is refused, naming the model.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Ilker ManapandClaude Opus 5.5 committed 2026-10-15 00:22:01 +02:00
1 parent bda988b78f
commit c22528c26e
9 files changed
+220 -1

No files matched your search

+66 -1
View File
@@ -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
}
+25
View File
@@ -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)
}
}
+2
View File
@@ -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",
+2
View File
@@ -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",