Files
mGPUmanager/internal/scheduler/evicting_test.go
mAi 468317e395 fix(scheduler): mark lazy consumers (Unload but no Load) as not-loaded at startup
Live deploy on mRock surfaced a Schritt 5 bug: comfyui was always
treated as preloaded at scheduler startup, which made ensureFits()
short-circuit on the very first /v1/image request — exactly the
scenario eviction is supposed to handle. mvoice was never picked as
a victim, ComfyUI then OOM'd loading FLUX on top of the still-resident
mvoice.

Fix: replace the blanket 'every consumer starts loaded' init with a
heuristic — initialLoaded(cons):

  - VRAMManaged (ollama): true. We never track/evict it; the consumer
    runs its own LRU.
  - Load+Unload both present (mvoice): true. Designed to be controllable;
    typically preloads in its own lifespan.
  - Unload only, no Load (comfyui): false. Lazy — FLUX isn't resident
    until the first /prompt, so we shouldn't bill its 13 GiB against the
    GPU budget until then.
  - SystemdUnit only (whisper-server): true. Always-on, model loaded at
    process start.
  - Empty: true. Safe fallback.

Verified live on mRock (2026-05-15):

  Before /v1/image:  nvidia-smi 8963 MiB used; mvoice gpu_resident_mib 2345
  POST /v1/image:    HTTP 400 from upstream (empty workflow), broker did
                     trigger eviction before forwarding
  After:             nvidia-smi 6547 MiB used; mvoice gpu_resident_mib 9
                     (~CUDA context only); scheduler.evictions = 2
  POST /v1/tts:      audio_url returned, tts_ms 670, audio 3.5 s
  After reload:      nvidia-smi 8943 MiB used; mvoice gpu_resident_mib 2917

Test: TestInitialLoadedHeuristic pins the four cases down so this
doesn't regress when someone adds a fifth consumer type.

Refs: m/mGPUmanager#1 (live deploy).
2026-05-15 16:54:11 +02:00

279 lines
9.4 KiB
Go

package scheduler
import (
"context"
"io"
"log/slog"
"net/http"
"net/http/httptest"
"sync/atomic"
"testing"
"time"
"mgit.msbls.de/m/mGPUmanager/internal/config"
"mgit.msbls.de/m/mGPUmanager/internal/gpu"
"mgit.msbls.de/m/mGPUmanager/internal/registry"
)
func silentLogger() *slog.Logger {
return slog.New(slog.NewTextHandler(io.Discard, &slog.HandlerOptions{Level: slog.LevelError}))
}
// gpuStub implements just enough of gpu.Poller's surface for the evicting
// scheduler. We use the real Poller type (no interface yet) by hand-loading
// a Sample via a tiny wrapper.
//
// In practice we set gpu.Poller's internal sample via NewPoller + a goroutine.
// For tests we sidestep that by using a real Poller with a fake nvidia-smi —
// but the simpler path is to construct a Poller, store a Sample, and skip
// Run. We do that by exposing a tiny helper here.
// makeGPU returns a Poller pre-loaded with the given free/total values.
// It never calls nvidia-smi.
func makeGPU(t *testing.T, freeMiB, totalMiB int) *gpu.Poller {
t.Helper()
p := gpu.NewPoller(time.Hour, silentLogger())
// gpu.Poller.Last() reads from an internal Sample. We can't poke it
// directly without exporting state, so we use a sub-test trick: run
// sampleOnce against a fake nvidia-smi command. But that needs a PATH
// override and is brittle. Instead, expose a SetForTest helper.
gpu.SetSampleForTest(p, freeMiB, totalMiB)
return p
}
// fakeConsumer hosts /api/admin/{load,unload} so the evicting scheduler can
// exercise the HTTP eviction path.
type fakeConsumer struct {
srv *httptest.Server
unloadHit atomic.Int32
loadHit atomic.Int32
}
func newFakeConsumer(t *testing.T) *fakeConsumer {
t.Helper()
fc := &fakeConsumer{}
mux := http.NewServeMux()
mux.HandleFunc("GET /api/health", func(w http.ResponseWriter, _ *http.Request) {
w.Header().Set("Content-Type", "application/json")
_, _ = w.Write([]byte(`{"loaded":true,"gpu_resident_mib":2800}`))
})
mux.HandleFunc("POST /api/admin/unload", func(w http.ResponseWriter, _ *http.Request) {
fc.unloadHit.Add(1)
w.WriteHeader(200)
})
mux.HandleFunc("POST /api/admin/load", func(w http.ResponseWriter, _ *http.Request) {
fc.loadHit.Add(1)
w.WriteHeader(200)
})
mux.HandleFunc("POST /prompt", func(w http.ResponseWriter, _ *http.Request) {
w.WriteHeader(200)
})
mux.HandleFunc("POST /api/free", func(w http.ResponseWriter, _ *http.Request) {
fc.unloadHit.Add(1)
w.WriteHeader(200)
})
fc.srv = httptest.NewServer(mux)
return fc
}
func buildCfg(mvoiceURL, comfyURL string) *config.Config {
return &config.Config{
Listen: "127.0.0.1:0",
GPU: config.GPU{TotalMiB: 16376, ReservedMiB: 1024, PollIntervalSeconds: 2},
Routing: map[config.EndpointKind]string{
config.KindTTS: "mvoice",
config.KindImage: "comfyui",
},
Consumers: map[string]*config.Consumer{
"mvoice": {
URL: mvoiceURL,
Health: config.Route{Method: "GET", Path: "/api/health"},
Paths: map[config.EndpointKind]config.Route{
config.KindTTS: {Method: "POST", Path: "/api/synthesize"},
},
VRAMResidentMiB: 2800,
Load: &config.Route{Method: "POST", Path: "/api/admin/load"},
Unload: &config.Route{Method: "POST", Path: "/api/admin/unload"},
CanCoexistWith: []string{"whisper-server", "ollama"},
Priority: 3,
MaxConcurrency: 1,
},
"comfyui": {
URL: comfyURL,
Health: config.Route{Method: "GET", Path: "/system_stats"},
Paths: map[config.EndpointKind]config.Route{
config.KindImage: {Method: "POST", Path: "/prompt"},
},
VRAMResidentMiB: 13000,
Unload: &config.Route{
Method: "POST",
Path: "/api/free",
Body: `{"unload_models":true,"free_memory":true}`,
},
CanCoexistWith: []string{},
Priority: 1,
MaxConcurrency: 1,
},
},
}
}
// TestInitialLoadedHeuristic pins the comfyui-isn't-preloaded rule down:
// a consumer with Unload but no Load is lazy; everything else is assumed
// resident at startup.
func TestInitialLoadedHeuristic(t *testing.T) {
cases := []struct {
name string
cons *config.Consumer
want bool
}{
{"vram_managed (ollama)", &config.Consumer{VRAMManaged: true}, true},
{"load+unload (mvoice)", &config.Consumer{
Load: &config.Route{Path: "/load"},
Unload: &config.Route{Path: "/unload"},
}, true},
{"unload only — lazy (comfyui)", &config.Consumer{
Unload: &config.Route{Path: "/api/free"},
}, false},
{"systemd unit only (whisper-server)", &config.Consumer{
SystemdUnit: "whisper-server.service",
}, true},
{"empty consumer", &config.Consumer{}, true},
}
for _, c := range cases {
t.Run(c.name, func(t *testing.T) {
if got := initialLoaded(c.cons); got != c.want {
t.Errorf("initialLoaded = %v, want %v", got, c.want)
}
})
}
}
// TestEvictingSkipsWhenAlreadyResident verifies the no-op fast path: a job
// for an already-loaded consumer with plenty of free VRAM runs without any
// unload call.
func TestEvictingSkipsWhenAlreadyResident(t *testing.T) {
mvoice := newFakeConsumer(t)
defer mvoice.srv.Close()
comfy := newFakeConsumer(t)
defer comfy.srv.Close()
cfg := buildCfg(mvoice.srv.URL, comfy.srv.URL)
reg := registry.New(cfg, silentLogger())
g := makeGPU(t, 8192, 16376) // plenty of headroom
e := NewEvicting(cfg, reg, g, silentLogger())
if err := e.Run(context.Background(), "mvoice", func(ctx context.Context) error { return nil }); err != nil {
t.Fatal(err)
}
if mvoice.unloadHit.Load() != 0 {
t.Errorf("unexpected unload hits on mvoice: %d", mvoice.unloadHit.Load())
}
if comfy.unloadHit.Load() != 0 {
t.Errorf("unexpected unload hits on comfyui: %d", comfy.unloadHit.Load())
}
}
// TestEvictingFreesNonCoexistentVictim simulates the canonical scenario from
// the design: a TTS request comes in while comfyui is hogging 13 GiB. mvoice
// is not coexistent with comfyui (per cfg), so the scheduler must call
// comfyui's /api/free before letting the TTS job run.
func TestEvictingFreesNonCoexistentVictim(t *testing.T) {
mvoice := newFakeConsumer(t)
defer mvoice.srv.Close()
comfy := newFakeConsumer(t)
defer comfy.srv.Close()
cfg := buildCfg(mvoice.srv.URL, comfy.srv.URL)
reg := registry.New(cfg, silentLogger())
// Only 1 GiB free — mvoice (2.8 GiB) won't fit until comfyui (13 GiB)
// is evicted.
g := makeGPU(t, 1024, 16376)
e := NewEvicting(cfg, reg, g, silentLogger())
// Force the believed-loaded state so eviction kicks in (Run treats
// 'already loaded' as a no-op fast path).
e.SetLoadedForTest("mvoice", false)
e.SetLoadedForTest("comfyui", true)
// After the eviction unload call lands, we want fits() to return true
// for the next iteration — patch the GPU sample to reflect the freed
// memory by swapping the poller before the second fits() check is hit.
// We accomplish that by stubbing the unload handler to also bump the
// sample.
comfy.srv.Config.Handler = withHook(comfy.srv.Config.Handler, func() {
gpu.SetSampleForTest(g, 14000, 16376)
})
if err := e.Run(context.Background(), "mvoice", func(ctx context.Context) error { return nil }); err != nil {
t.Fatal(err)
}
if got := comfy.unloadHit.Load(); got != 1 {
t.Errorf("comfyui unload hit count = %d, want 1", got)
}
if got := mvoice.loadHit.Load(); got != 1 {
t.Errorf("mvoice load hit count = %d, want 1", got)
}
if got := e.Stats().Evictions; got != 1 {
t.Errorf("stats.Evictions = %d, want 1", got)
}
}
// TestEvictingHonoursCoexistence ensures we never evict a consumer that the
// target declared compatible. mvoice can coexist with ollama, so ollama must
// not be picked even if it's the LRU candidate.
func TestEvictingHonoursCoexistence(t *testing.T) {
mvoice := newFakeConsumer(t)
defer mvoice.srv.Close()
comfy := newFakeConsumer(t)
defer comfy.srv.Close()
cfg := buildCfg(mvoice.srv.URL, comfy.srv.URL)
// Add a stub ollama with an unload endpoint, mark coexistent.
ollama := newFakeConsumer(t)
defer ollama.srv.Close()
cfg.Consumers["ollama"] = &config.Consumer{
URL: ollama.srv.URL,
Health: config.Route{Method: "GET", Path: "/api/health"},
Paths: map[config.EndpointKind]config.Route{},
VRAMResidentMiB: 2000,
Unload: &config.Route{Method: "POST", Path: "/api/admin/unload"},
CanCoexistWith: []string{"mvoice"},
MaxConcurrency: 1,
}
reg := registry.New(cfg, silentLogger())
g := makeGPU(t, 1000, 16376)
e := NewEvicting(cfg, reg, g, silentLogger())
e.SetLoadedForTest("mvoice", false)
e.SetLoadedForTest("comfyui", true)
e.SetLoadedForTest("ollama", true)
comfy.srv.Config.Handler = withHook(comfy.srv.Config.Handler, func() {
gpu.SetSampleForTest(g, 14000, 16376)
})
if err := e.Run(context.Background(), "mvoice", func(ctx context.Context) error { return nil }); err != nil {
t.Fatal(err)
}
if got := ollama.unloadHit.Load(); got != 0 {
t.Errorf("ollama (coexistent) unloaded %d times; should be 0", got)
}
if got := comfy.unloadHit.Load(); got != 1 {
t.Errorf("comfyui unload hit count = %d, want 1", got)
}
}
// ───── helpers ────────────────────────────────────────────────────────────
// withHook wraps an http.Handler so each call invokes hook() before
// delegating to the original handler. Used to simulate VRAM being freed
// the instant comfyui's /api/free returns.
func withHook(h http.Handler, hook func()) http.Handler {
return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
hook()
h.ServeHTTP(w, r)
})
}