A judge reserves two cores and 2 GB, with no prompt cache
CI / test (push) Successful in 9m1s

Checked by a single-machine install of v0.14.6 on a 4-core, 8 GB VM. The
second judge fitted nowhere: v0.14.6 reserved 4 GB a judge, a figure from
one replica that grew to 3.2 GB through llama.cpp's prompt cache (up to
8 GB by default), which a judge never reuses. Once it fitted, every
judgment timed out. v0.14.6 gave a judge one core below 8 CPUs, and a
judgment of about 280 tokens takes 8.0 s on one core, 4.2 s on two and
2.3 s on four, against a 5 s budget: 0 of 20 requests went through.

A judge now runs with --cache-ram 0, uses about 1.6 GB and reserves 2,
and reserves two cores (one only below four CPUs). On the same VM, 19 of
20 went through.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Ilker ManapandClaude Opus 5.5 committed 2026-10-16 19:22:33 +02:00
1 parent 0def6eaba1
commit f8206c3540
4 files changed
+48 -23

No files matched your search

+15
View File
@@ -24,6 +24,21 @@ is only released once its **exit criterion** is met, not when its code lands.
Until v1.0.0 the minor version is the breaking one: configuration formats and Go
APIs change between minors without a deprecation period.
## Unreleased
### Fixed
- **Both judges fit an 8 GB machine again.** v0.14.6 had `server init` reserve 4 GB a judge replica. That figure came
from one replica at 3.2 GB after a day of traffic. The growth was llama.cpp's prompt cache, up to 8 GB by default,
which a judge never reuses: each text it reads is different. On a 4-core, 8 GB machine the second judge then fitted
nowhere ("has 7 GB of memory with 1 free, needs 4"). The judge now runs with `--cache-ram 0`, uses about 1.6 GB and
stays there, and reserves 2.
- **A judge gets the two cores it needs.** v0.14.6 gave each judge replica one core on machines under eight. A judgment
of about 280 tokens takes 8.0 s on one core, 4.2 s on two and 2.3 s on four, against the 5 s budget, so on a 4-core
machine every judgment timed out and every request was refused. A judge replica now reserves two cores, one only
below four CPUs. On a 4-core machine the starter reserves none and shares what is left. A judge on two cores
stays close to its budget: give it more cores or a GPU, or raise `latency_budget`, where requests must not wait.
## v0.14.6 — 2026-10-16
Setup that says what each part of an install is and lets every one be
+5 -5
View File
@@ -183,12 +183,12 @@ func TestTheJudgesAndTheStarterShareTheCPUsRatherThanEachTakingAll(t *testing.T)
in serverInit
judge, starter int
}{
{serverInit{cpus: 4}, 1, 2},
{serverInit{cpus: 4}, 2, 0}, // two a judge, and the starter shares what is left
{serverInit{cpus: 8}, 2, 4},
{serverInit{cpus: 32}, 2, 8},
{serverInit{cpus: 2}, 1, 0}, // nothing left: the starter shares rather than fits nowhere
{serverInit{cpus: 2}, 1, 0}, // below four, one each, so both fit
{serverInit{cpus: 1}, 0, 0},
{serverInit{cpus: 4, noGuardModel: true}, 1, 4},
{serverInit{cpus: 4, noGuardModel: true}, 2, 4},
{serverInit{cpus: 8, judge: judgeSpec{Replicas: 3}}, 2, 2},
{serverInit{cpus: 4, noNode: true}, 2, 0}, // on nodes not known yet
} {
@@ -198,8 +198,8 @@ func TestTheJudgesAndTheStarterShareTheCPUsRatherThanEachTakingAll(t *testing.T)
}
}
m := guardModel("http://x/g.gguf", 2, 1)
if p := m.Profiles[0]; p.CPUCores != 1 || p.MinMemoryGB != 4 {
t.Errorf("judge profile: %d cores, %d GB", p.CPUCores, p.MinMemoryGB)
if p := m.Profiles[0]; p.CPUCores != 1 || p.MinMemoryGB != 2 || !slices.Contains(p.Args, "--cache-ram") {
t.Errorf("judge profile: %d cores, %d GB, args %v", p.CPUCores, p.MinMemoryGB, p.Args)
}
s := starterModel(starterChoices[0], defaultContext, 2, "http://x/q.gguf", []string{"6000-24g"}, true)
if cpu := s.Profiles[len(s.Profiles)-1]; cpu.Name != "cpu" || cpu.CPUCores != 2 || s.Profiles[0].CPUCores != 0 {
+27 -17
View File
@@ -496,10 +496,13 @@ func guardModel(source string, replicas, cores int) catalog.Model {
// judge that is not there refuses everything.
Scaling: catalog.Scaling{Min: replicas, Max: replicas},
Profiles: []catalog.Profile{{
// It uses about 3.2 GB with its context: reserving 2 filled nodes
// past their memory.
Name: "cpu", Engine: "llamacpp", Priority: 100, MinMemoryGB: 4, CPUCores: cores,
Args: []string{"--ctx-size", "8192"},
// About 1.6 GB with its context, which stays so with no prompt
// cache: llama.cpp keeps up to 8 GB of prompts by default, and a
// judge's, each a different text, are not asked again. With the
// cache one grew to 3.2 GB over a day of traffic; reserving 4 for
// that left a 7 GB machine room for only one of the two.
Name: "cpu", Engine: "llamacpp", Priority: 100, MinMemoryGB: judgeMemoryGB, CPUCores: cores,
Args: []string{"--ctx-size", "8192", "--cache-ram", "0"},
}},
}
}
@@ -716,14 +719,16 @@ type cpuSplit struct{ judge, starter int }
// the starter on one 4-core control plane all ran on all four: under load
// most judgments missed their budget and the requests were refused.
//
// When this machine is the node, by its own CPUs: a core a judge under 8,
// two from 8, and the starter what is left, up to 8, or no reservation at
// all when nothing is, rather than a starter that fits nowhere. Otherwise
// two a judge, on whichever node takes it, and the starter, on a node not
// known yet, none.
// A judge needs two cores: a judgment of about 280 tokens took 8.0 s on
// one, 4.2 s on two and 2.3 s on four, against a 5 s budget. When this
// machine is the node, each judge replica gets two (one each only below four
// CPUs, where two would not fit both), and the starter what is left, up to
// 8, or no reservation at all when nothing is, rather than a starter that
// fits nowhere. Otherwise two a judge, on whichever node takes it, and the
// starter, on a node not known yet, none.
func (in *serverInit) cpuPlan() cpuSplit {
if in.noNode {
return cpuSplit{judge: 2}
return cpuSplit{judge: judgeCores}
}
n := in.cpus
if n <= 0 {
@@ -736,13 +741,18 @@ func (in *serverInit) cpuPlan() cpuSplit {
if !in.noGuardModel {
judges = in.judgeFor().Replicas
}
judge := 1
if n >= 8 {
judge = 2
}
left := n - judges*judge
if left < 1 {
left = 0
judge := judgeCores
if judges > 0 && judges*judge > n {
judge = max(1, n/judges)
}
left := max(n-judges*judge, 0)
return cpuSplit{judge: judge, starter: min(left, 8)}
}
// judgeCores and judgeMemoryGB are what a judge replica needs: two cores
// for a judgment within its budget, and 2 GB for its 1.6 with no prompt
// cache.
const (
judgeCores = 2
judgeMemoryGB = 2
)
+1 -1
View File
@@ -282,7 +282,7 @@ func newInstallForm(in *serverInit, host string, found []string, ha bool) *insta
"The request is served unjudged and marked so in the logs. Nothing is refused for the judge's sake.",
}
f.judgeCopies = textSetupField("Copies", strconv.Itoa(j.Replicas), "2",
"Each takes about 1 GB and two CPU threads on whichever node has room.")
"Each reserves 2 GB and two CPU cores on whichever node has room: on one core a judgment takes longer than its budget.")
f.judgeBudget = textSetupField("Waits at most", durationYAML(j.Budget), "5s",
"A short text is judged in about two seconds on CPUs.")
judgeFields := []*setupField{f.judgeFail, f.judgeCopies, f.judgeBudget, headingField("Refuses")}