Files
llmgw/internal/agent/discovery/host.go
T
Ilker ManapandClaude Opus 5.5 7d97c6b74f Install a node's engines on the node, instead of trusting what is there
A node ran whatever llama-server somebody had once unpacked on it. Onboarding
checked an engine only when told its path, installed none, and a node with
nothing on it registered as ready. The one time that went wrong it went wrong
invisibly: the binary was there, executable, and could not start for want of
libgomp1, while the node reported healthy.

The control plane now says which engines every node has, at which version —
`engines:` in server.yaml, defaulting to the llama.cpp build this release was
checked against — and each agent converges on it. What only the node can
know, the node works out: the build for its processor and driver (CUDA with
its runtime when an NVIDIA driver is loaded, CPU otherwise), the OS packages
that build needs from its own package sources, and whether the result runs.
Every archive is checked against the sha256 GitHub publishes for it and
unpacked refusing any path or link that leaves the install directory. ldd
names a missing library outright; running `--version` proves the rest. Only
then is a directory marked installed, and it is checked again every ten
minutes, because an unrelated upgrade can remove a library.

An engine is reported only once installed and seen to run, or when the
agent's configuration names it on purpose — never because something happened
to be on PATH. The agent now says when it has probed, so an empty list means
"has none" rather than "too old to say": a node rebuilt from nothing is no
longer a candidate for every model at the moment it can run none of them.
Placement's refusal carries the node's own words, "installing: downloading
llama-b10991-bin-ubuntu-x64.tar.gz (16 MB)", and so do `ctl nodes` and the
cluster page, along with a warning for a CPU without AVX2 — the VM CPU type
that makes everything five times slower and nothing look wrong.

Agents also report their OS, architecture, vector extensions and NVIDIA
driver. Onboarding installs llmgw itself from the CI release for the node's
architecture, checked against the release's checksums, rather than copying
whatever binary the operator's machine had.

vLLM and SGLang are not installed yet: they are Python environments of
several gigabytes, not an archive, and still have to be named in the agent's
configuration.

Also fixes a race on the weights fingerprint, which two heartbeat loops touch
during a failover.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-09-28 10:59:59 +02:00

136 lines
3.4 KiB
Go

package discovery
import (
"bufio"
"os"
"runtime"
"strings"
)
// Host is what a node is, as far as installing software on it is concerned.
type Host struct {
// OSID, OSVersion and OSCodename come from /etc/os-release: "ubuntu",
// "24.04", "noble".
OSID string `json:"os_id,omitempty"`
OSVersion string `json:"os_version,omitempty"`
OSCodename string `json:"os_codename,omitempty"`
// Arch is Debian's spelling, "amd64" or "arm64", because that is what
// package names and release archives are keyed by.
Arch string `json:"arch,omitempty"`
// CPUFlags are the vector extensions inference cares about, of those this
// processor has.
CPUFlags []string `json:"cpu_flags,omitempty"`
// NVIDIADriver is the loaded driver's version, empty when there is none.
NVIDIADriver string `json:"nvidia_driver,omitempty"`
}
// HasCPUFlag reports whether the processor has an extension.
func (h Host) HasCPUFlag(f string) bool {
for _, have := range h.CPUFlags {
if have == f {
return true
}
}
return false
}
// interestingFlags are the extensions that decide which build of an engine
// runs well. A VM on QEMU's default CPU model has none of them — and still
// runs, five times slower, with nothing saying why.
var interestingFlags = []string{"avx", "avx2", "fma", "f16c", "avx512f", "avx512_bf16", "amx_tile", "asimd", "sve"}
// DiscoverHost reads what the machine is.
func DiscoverHost() Host {
h := hostFromOSRelease("/etc/os-release")
h.Arch = debianArch(runtime.GOARCH)
h.CPUFlags = cpuFlagsFrom("/proc/cpuinfo")
h.NVIDIADriver = nvidiaDriverFrom("/proc/driver/nvidia/version")
return h
}
func hostFromOSRelease(path string) Host {
var h Host
f, err := os.Open(path) //nolint:gosec // a fixed system path
if err != nil {
return h
}
defer f.Close()
sc := bufio.NewScanner(f)
for sc.Scan() {
k, v, ok := strings.Cut(sc.Text(), "=")
if !ok {
continue
}
v = strings.Trim(v, `"'`)
switch k {
case "ID":
h.OSID = v
case "VERSION_ID":
h.OSVersion = v
case "VERSION_CODENAME":
h.OSCodename = v
}
}
return h
}
func debianArch(goarch string) string {
switch goarch {
case "amd64", "arm64":
return goarch
case "386":
return "i386"
case "arm":
return "armhf"
}
return goarch
}
func cpuFlagsFrom(path string) []string {
b, err := os.ReadFile(path) //nolint:gosec // a fixed system path
if err != nil {
return nil
}
have := map[string]bool{}
for _, line := range strings.Split(string(b), "\n") {
k, v, ok := strings.Cut(line, ":")
if !ok {
continue
}
// "flags" on x86, "Features" on ARM. The first processor is enough:
// a machine does not mix instruction sets across its cores.
if k = strings.TrimSpace(k); k != "flags" && k != "Features" {
continue
}
for _, f := range strings.Fields(v) {
have[f] = true
}
break
}
var out []string
for _, f := range interestingFlags {
if have[f] {
out = append(out, f)
}
}
return out
}
// nvidiaDriverFrom reads the loaded kernel module's version, which exists
// exactly when a driver is loaded — unlike nvidia-smi, which can be installed
// on a machine whose module failed to load.
func nvidiaDriverFrom(path string) string {
b, err := os.ReadFile(path) //nolint:gosec // a fixed system path
if err != nil {
return ""
}
// "NVRM version: NVIDIA UNIX x86_64 Kernel Module 570.86.15 Thu ..."
fields := strings.Fields(string(b))
for i, f := range fields {
if f == "Module" && i+1 < len(fields) {
return fields[i+1]
}
}
return ""
}