Files
llmgw/internal/agent/discovery/weights_test.go
T
Ilker ManapandClaude Opus 5 c13c567cc5
CI / test (push) Successful in 1m32s
Know where the weights are, without owning the storage
Storage will be different at every installation — Lustre, NFS over RDMA, an
image exported with SRP or NVMe-oF, local NVMe, an rsync — so llmgw observes
it rather than managing it. An agent told which directories to look in
reports what it can see; how the files got there is nobody's business here,
and a node told nothing reports nothing and behaves exactly as every node
did before this existed.

That fact buys two things. Placement prefers a node that already holds the
model over one that must first move tens of gigabytes, as a preference and
never a filter: fetching them is a valid answer, and refusing to place would
turn a slow start into no start at all. And ctl status shows what each
fabric holds.

Per fabric because fabrics are not connected to one another. Weights in
ib-a are not weights in ib-b — the only path between them is the management
network — so a model seeded into one fabric leaves the rest of the cluster
unable to serve it without a slow copy, and nothing said so before now. One
shared filer counts the same as twenty local copies, because "can this
fabric serve it" is the question being asked.

Three distinctions the tests hold to, because each is a way to get this
quietly wrong. A node that was never asked to look must rank with one that
looked and has nothing, or enabling the feature on a single node would
drain every other. An empty directory is not weights, because a
half-finished download looks exactly like one and would send placement to a
node whose engine then fails to load. And local storage is never mistaken
for shared: a filer read as local costs an unnecessary copy, while local
read as shared would have the cluster believe a model is everywhere when it
is on one node.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-09-17 08:21:16 +02:00

115 lines
3.7 KiB
Go

package discovery
import (
"os"
"path/filepath"
"testing"
)
func write(t *testing.T, path string, size int) {
t.Helper()
if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
t.Fatal(err)
}
if err := os.WriteFile(path, make([]byte, size), 0o644); err != nil {
t.Fatal(err)
}
}
func TestScanFindsFilesAndDirectoriesOfWeights(t *testing.T) {
dir := t.TempDir()
write(t, filepath.Join(dir, "qwen-3b-q4.gguf"), 1024)
// A HuggingFace checkout: a directory of shards rather than one file.
write(t, filepath.Join(dir, "Llama-3.3-70B", "model-00001.safetensors"), 2048)
write(t, filepath.Join(dir, "Llama-3.3-70B", "config.json"), 64)
got := ScanWeights([]string{dir})
if len(got) != 2 {
t.Fatalf("found %d entries: %+v", len(got), got)
}
byName := map[string]int64{}
for _, w := range got {
byName[filepath.Base(w.Path)] = w.SizeBytes
}
if byName["qwen-3b-q4.gguf"] != 1024 {
t.Errorf("gguf size = %d", byName["qwen-3b-q4.gguf"])
}
// The directory's size is everything in it, config included: what matters
// is how much would have to be moved to put it somewhere else.
if byName["Llama-3.3-70B"] != 2048+64 {
t.Errorf("checkout size = %d, want %d", byName["Llama-3.3-70B"], 2048+64)
}
}
func TestAnEmptyDirectoryIsNotWeights(t *testing.T) {
// A half-finished download looks exactly like this, and reporting it as
// present would have placement choose a node whose engine then fails to
// load — the failure this whole feature exists to avoid.
dir := t.TempDir()
if err := os.MkdirAll(filepath.Join(dir, "downloading"), 0o755); err != nil {
t.Fatal(err)
}
write(t, filepath.Join(dir, "downloading", "README.md"), 10)
if got := ScanWeights([]string{dir}); len(got) != 0 {
t.Errorf("a directory with no weight files was reported: %+v", got)
}
}
func TestNotesAndScriptsAreNotReportedAsModels(t *testing.T) {
dir := t.TempDir()
write(t, filepath.Join(dir, "notes.txt"), 10)
write(t, filepath.Join(dir, "fetch.sh"), 10)
write(t, filepath.Join(dir, "real.gguf"), 10)
got := ScanWeights([]string{dir})
if len(got) != 1 || filepath.Base(got[0].Path) != "real.gguf" {
t.Errorf("scan returned %+v", got)
}
}
func TestAMissingDirectoryIsNotAnError(t *testing.T) {
// A mount that has not come back yet looks exactly like a directory that is
// not there, and the node is simply one without those weights until it is.
got := ScanWeights([]string{filepath.Join(t.TempDir(), "never-existed")})
if len(got) != 0 {
t.Errorf("got %+v", got)
}
}
func TestScanningNothingIsTheDefaultAndConstrainsNothing(t *testing.T) {
// What makes the feature optional: an agent with no directories configured
// behaves exactly as every agent did before this existed.
if got := ScanWeights(nil); len(got) != 0 {
t.Errorf("got %+v", got)
}
if got := ScanWeights([]string{""}); len(got) != 0 {
t.Errorf("an empty directory name produced %+v", got)
}
}
func TestTheSamePathIsReportedOnce(t *testing.T) {
// Two configured directories can overlap, or be the same mount reached two
// ways. Counting a model twice would overstate what a fabric holds.
dir := t.TempDir()
write(t, filepath.Join(dir, "m.gguf"), 10)
if got := ScanWeights([]string{dir, dir}); len(got) != 1 {
t.Errorf("the same directory twice produced %+v", got)
}
}
func TestLocalStorageIsNeverMistakenForShared(t *testing.T) {
// The conservative direction. A shared filer read as local costs an
// unnecessary copy; local read as shared would have the cluster believe a
// model is everywhere when it is on one node.
dir := t.TempDir()
write(t, filepath.Join(dir, "m.gguf"), 10)
got := ScanWeights([]string{dir})
if len(got) != 1 {
t.Fatalf("got %+v", got)
}
if got[0].Shared {
t.Error("a temporary directory on local disk was reported as shared")
}
}