Auto-unload idle models and evict on model switch across all workers

Every ASR/TTS worker (faster-whisper, qwen3, vibevoice, xtts) kept every
model it ever loaded resident in GPU memory forever, and switching to a
different model (e.g. a different whisper size) just added another one
alongside it rather than freeing the old one. Combined with the shared
24GB GPU, this made memory pressure only ever go up.

Now: only one model stays resident per worker at a time (loading a
different model_id evicts the previous one first), and the whole cache
(plus, for faster-whisper, the diarization pipeline) is dropped after 2
minutes of no requests. Verified end-to-end: the idle timer actually fires
and reclaims memory, and switching qwen3 models evicts the old one.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

app/asr/workers/faster_whisper_worker.py

@@ -1,7 +1,10 @@
 from __future__ import annotations
 
 import argparse
+import asyncio
+import gc
 import tempfile
+import time
 from pathlib import Path
 from typing import Any, Dict, List, Optional, Union
 
@@ -32,10 +35,55 @@ app = FastAPI(title="ASR Faster-Whisper Worker")
 _MODEL_CACHE: Dict[str, WhisperModel] = {}
 _DIARIZATION_PIPELINE: Any = None
 
+# Idle-unload / model-switch eviction: only one whisper model stays resident
+# at a time (switching sizes frees the old one), and everything (including
+# the diarization pipeline) is dropped after IDLE_UNLOAD_SECONDS of no
+# requests, so this backend doesn't permanently hog GPU memory shared with
+# the other ASR workers.
+IDLE_UNLOAD_SECONDS = 120
+_last_used: float = 0.0
+_active_requests: int = 0
+
+
+def _free_gpu() -> None:
+    gc.collect()
+    try:
+        import torch
+        if DEVICE == "cuda" and torch.cuda.is_available():
+            torch.cuda.empty_cache()
+    except Exception:
+        pass
+
+
+def _unload_whisper_models() -> None:
+    if not _MODEL_CACHE:
+        return
+    print(f"[faster_whisper] unloading {list(_MODEL_CACHE.keys())}", flush=True)
+    _MODEL_CACHE.clear()
+    _free_gpu()
+
+
+def _unload_diarization() -> None:
+    global _DIARIZATION_PIPELINE
+    if _DIARIZATION_PIPELINE is None:
+        return
+    print("[faster_whisper] unloading diarization pipeline", flush=True)
+    _DIARIZATION_PIPELINE = None
+    _free_gpu()
+
+
+async def _idle_unload_loop() -> None:
+    while True:
+        await asyncio.sleep(30)
+        if _active_requests == 0 and _last_used and (time.monotonic() - _last_used) >= IDLE_UNLOAD_SECONDS:
+            _unload_whisper_models()
+            _unload_diarization()
+
 
 @app.on_event("startup")
-def startup() -> None:
+async def startup() -> None:
     ensure_runtime_dirs()
+    asyncio.create_task(_idle_unload_loop())
 
 
 @app.get("/health")
@@ -75,6 +123,8 @@ async def transcribe(
         tmp.write(await file.read())
         tmp_path = Path(tmp.name)
 
+    global _active_requests, _last_used
+    _active_requests += 1
     try:
         resolved_model = resolve_custom_model_path(custom_model_path) or model
         result = _transcribe(
@@ -107,10 +157,16 @@ async def transcribe(
         raise HTTPException(status_code=500, detail=f"Worker failure: {exc}") from exc
     finally:
         tmp_path.unlink(missing_ok=True)
+        _free_gpu()
+        _active_requests -= 1
+        _last_used = time.monotonic()
 
 
 def _load_model(model_name: str) -> WhisperModel:
     if model_name not in _MODEL_CACHE:
+        if _MODEL_CACHE:
+            # Only one whisper model resident at a time — switching sizes frees the old one.
+            _unload_whisper_models()
         _MODEL_CACHE[model_name] = WhisperModel(
             model_name,
             device=DEVICE,

app/asr/workers/qwen3_worker.py

@@ -1,9 +1,11 @@
 from __future__ import annotations
 
 import argparse
+import asyncio
 import gc
 import subprocess
 import tempfile
+import time
 from pathlib import Path
 from typing import Any, Dict, Optional, Tuple
 
@@ -29,6 +31,14 @@ app = FastAPI(title="ASR Qwen3 Worker")
 # and drive it through Qwen3ASRProcessor.apply_transcription_request().
 _MODEL_CACHE: Dict[str, Tuple[Any, Any]] = {}
 
+# Idle-unload / model-switch eviction: only one model stays resident at a
+# time, and the whole cache is dropped after IDLE_UNLOAD_SECONDS of no
+# requests, so this backend doesn't permanently hog GPU memory shared with
+# the other ASR workers.
+IDLE_UNLOAD_SECONDS = 120
+_last_used: float = 0.0
+_active_requests: int = 0
+
 LANG_MAP = {
     "ko": "korean", "en": "english", "ja": "japanese", "zh": "chinese",
     "fr": "french", "de": "german", "es": "spanish", "ru": "russian",
@@ -70,10 +80,29 @@ def _to_wav(src_path: Path) -> Path:
     return wav_path
 
 
+def _unload_models() -> None:
+    if not _MODEL_CACHE:
+        return
+    _log(f"unloading {list(_MODEL_CACHE.keys())}")
+    _MODEL_CACHE.clear()
+    _free_gpu()
+
+
+async def _idle_unload_loop() -> None:
+    while True:
+        await asyncio.sleep(30)
+        if _active_requests == 0 and _last_used and (time.monotonic() - _last_used) >= IDLE_UNLOAD_SECONDS:
+            _unload_models()
+
+
 def _load_model(model_id: str) -> Tuple[Any, Any]:
     if model_id in _MODEL_CACHE:
         return _MODEL_CACHE[model_id]
 
+    if _MODEL_CACHE:
+        # Only one model resident at a time — switching models frees the old one.
+        _unload_models()
+
     import torch
     from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
 
@@ -96,8 +125,9 @@ def _load_model(model_id: str) -> Tuple[Any, Any]:
 
 
 @app.on_event("startup")
-def startup() -> None:
+async def startup() -> None:
     ensure_runtime_dirs()
+    asyncio.create_task(_idle_unload_loop())
 
 
 @app.get("/health")
@@ -118,6 +148,8 @@ async def transcribe(
         tmp.write(await file.read())
         tmp_path = Path(tmp.name)
 
+    global _active_requests, _last_used
+    _active_requests += 1
     wav_path: Optional[Path] = None
     try:
         import torch
@@ -162,6 +194,21 @@ async def transcribe(
         tmp_path.unlink(missing_ok=True)
         if wav_path is not None:
             wav_path.unlink(missing_ok=True)
+        # Long files blow up KV-cache/activation memory; without this, that
+        # memory stays reserved by this process and starves the other backends
+        # sharing the same GPU until the container restarts. (del locals()[...]
+        # does NOT work in CPython, hence the explicit names.)
+        try:
+            del inputs
+        except NameError:
+            pass
+        try:
+            del generated
+        except NameError:
+            pass
+        _free_gpu()
+        _active_requests -= 1
+        _last_used = time.monotonic()
 
 
 if __name__ == "__main__":

app/asr/workers/vibevoice_worker.py

@@ -1,8 +1,10 @@
 from __future__ import annotations
 
 import argparse
+import asyncio
 import gc
 import tempfile
+import time
 from pathlib import Path
 from typing import Any, Dict, List, Optional
 
@@ -24,15 +26,52 @@ app = FastAPI(title="ASR VibeVoice Worker")
 
 _MODEL_CACHE: Dict[str, Any] = {}
 
+# Idle-unload / model-switch eviction: only one model stays resident at a
+# time, and the whole cache is dropped after IDLE_UNLOAD_SECONDS of no
+# requests, so this backend doesn't permanently hog GPU memory shared with
+# the other ASR workers.
+IDLE_UNLOAD_SECONDS = 120
+_last_used: float = 0.0
+_active_requests: int = 0
+
 
 def _log(msg: str) -> None:
     print(f"[vibevoice] {msg}", flush=True)
 
 
+def _free_gpu() -> None:
+    gc.collect()
+    try:
+        import torch
+        if DEVICE == "cuda" and torch.cuda.is_available():
+            torch.cuda.empty_cache()
+    except Exception:
+        pass
+
+
+def _unload_models() -> None:
+    if not _MODEL_CACHE:
+        return
+    _log(f"unloading {list(_MODEL_CACHE.keys())}")
+    _MODEL_CACHE.clear()
+    _free_gpu()
+
+
+async def _idle_unload_loop() -> None:
+    while True:
+        await asyncio.sleep(30)
+        if _active_requests == 0 and _last_used and (time.monotonic() - _last_used) >= IDLE_UNLOAD_SECONDS:
+            _unload_models()
+
+
 def _load_model(model_id: str) -> Any:
     if model_id in _MODEL_CACHE:
         return _MODEL_CACHE[model_id]
 
+    if _MODEL_CACHE:
+        # Only one model resident at a time — switching models frees the old one.
+        _unload_models()
+
     import torch
     from vibevoice.modular.modeling_vibevoice_asr import VibeVoiceASRForConditionalGeneration
     from vibevoice.processor.vibevoice_asr_processor import VibeVoiceASRProcessor
@@ -59,8 +98,9 @@ def _load_model(model_id: str) -> Any:
 
 
 @app.on_event("startup")
-def startup() -> None:
+async def startup() -> None:
     ensure_runtime_dirs()
+    asyncio.create_task(_idle_unload_loop())
 
 
 @app.get("/health")
@@ -81,6 +121,8 @@ async def transcribe(
         tmp.write(await file.read())
         tmp_path = Path(tmp.name)
 
+    global _active_requests, _last_used
+    _active_requests += 1
     try:
         import torch
 
@@ -153,6 +195,25 @@ async def transcribe(
         raise HTTPException(status_code=500, detail=f"{type(e).__name__}: {e}")
     finally:
         tmp_path.unlink(missing_ok=True)
+        # Long files blow up KV-cache/activation memory; without this, that
+        # memory stays reserved by this process and starves the other backends
+        # sharing the same GPU until the container restarts. (del locals()[...]
+        # does NOT work in CPython, hence the explicit names.)
+        try:
+            del inputs
+        except NameError:
+            pass
+        try:
+            del output_ids
+        except NameError:
+            pass
+        try:
+            del generated_ids
+        except NameError:
+            pass
+        _free_gpu()
+        _active_requests -= 1
+        _last_used = time.monotonic()
 
 
 if __name__ == "__main__":

app/tts/workers/xtts_worker.py

@@ -1,8 +1,11 @@
 from __future__ import annotations
 
 import argparse
+import asyncio
+import gc
 import os
 import tempfile
+import time
 from pathlib import Path
 from typing import Any, Dict
 
@@ -28,6 +31,13 @@ app = FastAPI(title="TTS XTTS Worker")
 MODEL_NAME = "tts_models/multilingual/multi-dataset/xtts_v2"
 _MODEL_CACHE: Dict[str, Any] = {}
 
+# Idle-unload: the model is dropped after IDLE_UNLOAD_SECONDS of no synthesis
+# requests, so this backend doesn't permanently hog GPU memory shared with
+# the ASR workers.
+IDLE_UNLOAD_SECONDS = 120
+_last_used: float = 0.0
+_active_requests: int = 0
+
 
 def _log(msg: str) -> None:
     print(f"[xtts] {msg}", flush=True)
@@ -41,6 +51,31 @@ def _device() -> str:
         return "cpu"
 
 
+def _free_gpu() -> None:
+    gc.collect()
+    try:
+        import torch
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+    except Exception:
+        pass
+
+
+def _unload_model() -> None:
+    if not _MODEL_CACHE:
+        return
+    _log("unloading model (idle)")
+    _MODEL_CACHE.clear()
+    _free_gpu()
+
+
+async def _idle_unload_loop() -> None:
+    while True:
+        await asyncio.sleep(30)
+        if _active_requests == 0 and _last_used and (time.monotonic() - _last_used) >= IDLE_UNLOAD_SECONDS:
+            _unload_model()
+
+
 def _load_model() -> Any:
     if MODEL_NAME in _MODEL_CACHE:
         return _MODEL_CACHE[MODEL_NAME]
@@ -55,8 +90,9 @@ def _load_model() -> Any:
 
 
 @app.on_event("startup")
-def startup() -> None:
+async def startup() -> None:
     ensure_runtime_dirs()
+    asyncio.create_task(_idle_unload_loop())
 
 
 @app.get("/health")
@@ -74,10 +110,14 @@ def synthesize(
     if not speaker_path.exists():
         raise HTTPException(status_code=400, detail=f"speaker_wav not found: {speaker_wav}")
 
+    global _active_requests, _last_used
+    _active_requests += 1
     try:
         tts = _load_model()
     except Exception as e:
         _log(f"model load error: {type(e).__name__}: {e}")
+        _active_requests -= 1
+        _last_used = time.monotonic()
         raise HTTPException(status_code=500, detail=f"모델 로드 실패: {type(e).__name__}: {e}")
 
     with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
@@ -105,6 +145,9 @@ def synthesize(
         raise HTTPException(status_code=500, detail=f"{type(e).__name__}: {e}")
     finally:
         out_path.unlink(missing_ok=True)
+        _free_gpu()
+        _active_requests -= 1
+        _last_used = time.monotonic()
 
 
 if __name__ == "__main__":