fix(openai-compat): reuse cached engine instances in _resolve_engine

CHANGELOG.md

@@ -33,6 +33,7 @@ the frozen-backend fallback mirror it for their toolchains.
 - The OmniVoice guide now covers combining style attributes with a reference clip (consistent instruct stabilizes cloning; the reference wins conflicts), inline pronunciation control (pinyin / CMU phonemes), and corrects the claim that the default engine can't do voice design — it can, from attributes (#1565)
 
 ### Fixed
+- The OpenAI-compatible `/v1/audio/speech` route now reuses the shared cached engine for explicit `model` ids instead of constructing a fresh engine — and its sidecar/model load, a ~28s floor per call for subprocess engines — on every request, with the same single-engine-resident discipline `/generate` applies (#1614) — thanks @paoloantinori!
 - The setup wizard's RAM check no longer blocks 8 GB machines whose OS reports ~7.8 GB usable — the thresholds now tolerate reserved memory, and `OMNIVOICE_RAM_PREFLIGHT=0` turns a genuine block into a warning for those who accept the OOM risk (#1618)
 - Invisible watermarking now runs eagerly instead of through `torch.compile` — AudioSeal's lazy compile sent the first embed of every session into Inductor's C++ codegen, which failed outright on macOS hosts whose toolchain couldn't serve it and shipped the audio unmarked after a 30-40s wait; first embed drops from 9.70s to 0.26s (#1615) — thanks @paoloantinori!
 - The macOS Accessibility blocker now rechecks while visible and closes as soon as the grant is enabled instead of keeping a stale permission prompt on screen (#1609)

backend/api/routers/openai_compat.py

@@ -160,7 +160,9 @@ class VerboseTranscriptionResponse(BaseModel):
 
 def _resolve_engine(model_id: str):
     """Map an OpenAI model name to a VoiceStudio backend."""
-    from services.tts_backend import get_backend_class, get_active_tts_backend
+    from services.tts_backend import (
+        get_backend_class, get_active_tts_backend, get_engine_instance_for,
+    )
 
     # Accept OpenAI model names as pass-through to the active engine.
     if model_id in ("tts-1", "tts-1-hd"):
@@ -177,8 +179,18 @@ def _resolve_engine(model_id: str):
             )
         from services.tts_backend import OmniVoiceBackend
         if cls is OmniVoiceBackend:
+            # OmniVoice only ever runs as the shared active engine — the
+            # explicit-omnivoice request is the active-engine request.
             return get_active_tts_backend()
-        return cls()
+        # Cached singleton, not a fresh cls(): SubprocessBackend engines would
+        # spawn a sidecar process and reload their model on EVERY request, and
+        # register a new atexit hook each time (get_engine_instance's contract).
+        # No router-local cache on top of it: the shared cache is keyed by
+        # CLASS precisely so id rebinds/evictions can't serve a stale instance,
+        # and cross-engine memory discipline is create_speech's
+        # evict_other_tts_engines call (the same seam /generate uses) — not a
+        # bespoke unload here.
+        return get_engine_instance_for(model_id)
     except ValueError:
         raise HTTPException(
             status_code=400,
@@ -388,6 +400,15 @@ async def create_speech(req: SpeechRequest):
     # VRAM eviction runs in get_model()'s warm-return path now, covering every
     # native TTS generate (this route, WS TTS, dub, batch, audiobook).
 
+    # Single-active-engine memory discipline (MM2-01), the same call /generate
+    # makes before its load: hand back every OTHER resident TTS engine's model
+    # before this one warms up, so switching `model` ids across requests —
+    # explicit id → explicit id, or explicit id → the tts-1/omnivoice aliases —
+    # can't stack multi-GB engines/sidecars. No-op when nothing else is
+    # resident; opt out with OMNIVOICE_SINGLE_ENGINE_RESIDENT=0.
+    from services.engine_memory import evict_other_tts_engines
+    await evict_other_tts_engines(backend.id)
+
     # ── #1033/#1037/#1014: warm the engine under the LOAD budget before the
     # generate clock starts. The T4 verification (#1014) measured a fresh
     # install's first /v1/audio/speech burning its whole 300s generate budget

tests/test_openai_speech_engine_cache.py

@@ -0,0 +1,142 @@
+"""_resolve_engine must reuse engine instances across /v1/audio/speech requests.
+
+Regression: the direct engine-ID path did ``return cls()`` per request, so every
+``model: "pockettts"`` (or any SubprocessBackend engine) spawned a fresh sidecar
+process, re-imported torch and reloaded the engine's model — a ~28s floor per
+request on real hardware — and registered another atexit hook each time. The
+cached-singleton seam (``get_engine_instance_for``) exists precisely for this;
+the route just wasn't using it for explicit engine IDs (only tts-1/tts-1-hd got
+the shared active-engine instance).
+
+The flip side of caching is accumulation: cached explicit engines must not
+stack multi-GB residents when requests switch ``model`` ids. That is NOT a
+router-local unload cache (an id-keyed instance ref goes stale against the
+class-keyed shared cache — registry rebinds, idle sweeps) — the route calls
+``evict_other_tts_engines`` before warming the engine, the exact seam
+/generate uses (single-engine-resident policy, MM2-01), pinned here at the
+route level.
+"""
+from __future__ import annotations
+
+import os
+
+os.environ.setdefault("OMNIVOICE_MODEL", "test")
+os.environ.setdefault("OMNIVOICE_DISABLE_FILE_LOG", "1")
+
+import pytest
+
+
+def _tts_mod():
+    import importlib
+
+    return importlib.import_module("services.tts_backend")
+
+
+@pytest.fixture()
+def oc():
+    import importlib
+
+    return importlib.import_module("api.routers.openai_compat")
+
+
+def test_explicit_engine_id_resolves_to_the_cached_singleton(oc, monkeypatch):
+    svc = _tts_mod()
+
+    class _FakeBackend:
+        instances = 0
+
+        def __init__(self):
+            _FakeBackend.instances += 1
+
+        @staticmethod
+        def is_available():
+            return True, "ok"
+
+    monkeypatch.setattr(svc, "get_backend_class", lambda _id: _FakeBackend)
+
+    first = oc._resolve_engine("pockettts")
+    second = oc._resolve_engine("pockettts")
+    assert first is second
+    # Exactly one construction across both resolves: the cached singleton did
+    # the work, not a fresh cls() per request.
+    assert _FakeBackend.instances == 1
+
+
+def test_unknown_engine_id_still_400s(oc, monkeypatch):
+    from fastapi import HTTPException
+
+    svc = _tts_mod()
+
+    def _unknown(_id):
+        raise ValueError("no such engine")
+
+    monkeypatch.setattr(svc, "get_backend_class", _unknown)
+    with pytest.raises(HTTPException) as exc:
+        oc._resolve_engine("not-an-engine")
+    assert exc.value.status_code == 400
+    assert "Unknown model" in exc.value.detail
+
+
+# ── Cross-request memory discipline ─────────────────────────────────────────
+
+
+def _make_engine(tb, eid: str):
+    """A registry-real fake engine (same harness shape as
+    tests/test_text_normalization_routes.py) that counts its unloads."""
+    import torch
+
+    class _E(tb.TTSBackend):
+        id = eid
+        display_name = f"{eid} (test)"
+        supports_cloning = True
+        gpu_compat = ("cpu",)
+        unloads = 0
+
+        @property
+        def sample_rate(self) -> int:
+            return 24000
+
+        @property
+        def supported_languages(self) -> list[str]:
+            return ["multi"]
+
+        @classmethod
+        def is_available(cls):
+            return True, "ready"
+
+        def generate(self, text, **kw) -> torch.Tensor:
+            return torch.zeros(1, 24000)
+
+        def unload(self):
+            type(self).unloads += 1
+
+    return _E
+
+
+def test_speech_request_evicts_other_resident_engines(monkeypatch):
+    """Switching explicit `model` ids across requests must hand back the
+    outgoing engine's model (single-engine-resident policy) — the cached
+    singletons cannot accumulate residents."""
+    svc = _tts_mod()
+    a = _make_engine(svc, "fake-cache-a")
+    b = _make_engine(svc, "fake-cache-b")
+    monkeypatch.setitem(svc._REGISTRY, "fake-cache-a", a)
+    monkeypatch.setitem(svc._REGISTRY, "fake-cache-b", b)
+
+    from fastapi.testclient import TestClient
+    from main import app
+
+    client = TestClient(app, client=("127.0.0.1", 50000))
+
+    r1 = client.post("/v1/audio/speech", json={
+        "model": "fake-cache-a", "input": "hi", "response_format": "wav",
+    })
+    assert r1.status_code == 200, r1.text
+    assert a.unloads == 0  # the engine that just ran stays warm
+
+    r2 = client.post("/v1/audio/speech", json={
+        "model": "fake-cache-b", "input": "hi", "response_format": "wav",
+    })
+    assert r2.status_code == 200, r2.text
+    assert a.unloads == 1  # outgoing engine handed its model back
+    assert b.unloads == 0  # incoming engine untouched