Self-restart worker process when model loading fails

Discovered vibevoice sitting at 3.2GB resident GPU memory with
loaded_models: [] and no idle-unload ever firing for it again. Root cause:
a load attempt had OOM'd partway through (competing with an unrelated
ollama process on the same GPU), so _MODEL_CACHE was never populated —
the idle-unload loop only clears that cache, so it had nothing to act on,
even though the partially-constructed model had already left memory
allocated. gc.collect()+empty_cache() don't reliably reclaim memory from
an interrupted from_pretrained() call.

Confirmed a plain process restart does fully reclaim it, so each worker
now treats any load failure as fatal: log it and os._exit(1), letting
supervisord's autorestart=true respawn a clean process immediately.
Verified with a bogus model id — worker exits, respawns, and passes the
smoke test right after.

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(
@@ -107,16 +157,33 @@ 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:
-        _MODEL_CACHE[model_name] = WhisperModel(
-            model_name,
-            device=DEVICE,
-            compute_type=COMPUTE_TYPE,
-            download_root=str(MODEL_CACHE),
-        )
+        if _MODEL_CACHE:
+            # Only one whisper model resident at a time — switching sizes frees the old one.
+            _unload_whisper_models()
+        try:
+            _MODEL_CACHE[model_name] = WhisperModel(
+                model_name,
+                device=DEVICE,
+                compute_type=COMPUTE_TYPE,
+                download_root=str(MODEL_CACHE),
+            )
+        except Exception as e:
+            # A load interrupted partway (e.g. OOM) can leave CUDA memory
+            # fragmented/leaked in ways gc.collect()+empty_cache() don't
+            # reliably reclaim, and since _MODEL_CACHE never got populated
+            # the idle-unload loop has nothing to clean up either.
+            # Restarting the whole process is the only guaranteed way to get
+            # that memory back — supervisord's autorestart=true respawns it
+            # immediately.
+            print(f"[faster_whisper] model load failed, restarting process to reclaim GPU memory: {type(e).__name__}: {e}", flush=True)
+            os._exit(1)
     return _MODEL_CACHE[model_name]
 
 

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",
@@ -70,25 +80,54 @@ 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
 
     _log(f"loading model {model_id}")
     dtype = torch.float16 if DEVICE == "cuda" else torch.float32
 
-    processor = AutoProcessor.from_pretrained(model_id, cache_dir=str(MODEL_CACHE))
-    model = AutoModelForSpeechSeq2Seq.from_pretrained(
-        model_id,
-        dtype=dtype,
-        low_cpu_mem_usage=True,
-        cache_dir=str(MODEL_CACHE),
-    )
-    model.to(DEVICE)
-    model.eval()
+    try:
+        processor = AutoProcessor.from_pretrained(model_id, cache_dir=str(MODEL_CACHE))
+        model = AutoModelForSpeechSeq2Seq.from_pretrained(
+            model_id,
+            dtype=dtype,
+            low_cpu_mem_usage=True,
+            cache_dir=str(MODEL_CACHE),
+        )
+        model.to(DEVICE)
+        model.eval()
+    except Exception as e:
+        # A load interrupted partway (e.g. OOM) can leave CUDA memory
+        # fragmented/leaked in ways gc.collect()+empty_cache() don't reliably
+        # reclaim, and since _MODEL_CACHE never got populated the idle-unload
+        # loop has nothing to clean up either. Restarting the whole process
+        # is the only guaranteed way to get that memory back — supervisord's
+        # autorestart=true respawns it immediately.
+        _log(f"model load failed, restarting process to reclaim GPU memory: {type(e).__name__}: {e}")
+        os._exit(1)
 
     _MODEL_CACHE[model_id] = (model, processor)
     _log(f"model {model_id} loaded")
@@ -96,8 +135,9 @@ 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 +158,8 @@ 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 +204,21 @@ 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
 
@@ -24,34 +26,81 @@ 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
 
     _log(f"loading model {model_id}")
-    processor = VibeVoiceASRProcessor.from_pretrained(
-        model_id,
-        language_model_pretrained_name="Qwen/Qwen2.5-1.5B",
-        cache_dir=str(MODEL_CACHE),
-    )
-    model = VibeVoiceASRForConditionalGeneration.from_pretrained(
-        model_id,
-        dtype=torch.bfloat16 if DEVICE == "cuda" else torch.float32,
-        attn_implementation="sdpa",
-        trust_remote_code=True,
-        cache_dir=str(MODEL_CACHE),
-    )
-    model.to(DEVICE)
-    model.eval()
+    try:
+        processor = VibeVoiceASRProcessor.from_pretrained(
+            model_id,
+            language_model_pretrained_name="Qwen/Qwen2.5-1.5B",
+            cache_dir=str(MODEL_CACHE),
+        )
+        model = VibeVoiceASRForConditionalGeneration.from_pretrained(
+            model_id,
+            dtype=torch.bfloat16 if DEVICE == "cuda" else torch.float32,
+            attn_implementation="sdpa",
+            trust_remote_code=True,
+            cache_dir=str(MODEL_CACHE),
+        )
+        model.to(DEVICE)
+        model.eval()
+    except Exception as e:
+        # A load interrupted partway (e.g. OOM) can leave CUDA memory
+        # fragmented/leaked in ways gc.collect()+empty_cache() don't reliably
+        # reclaim, and since _MODEL_CACHE never got populated the idle-unload
+        # loop has nothing to clean up either. Restarting the whole process
+        # is the only guaranteed way to get that memory back — supervisord's
+        # autorestart=true respawns it immediately.
+        _log(f"model load failed, restarting process to reclaim GPU memory: {type(e).__name__}: {e}")
+        os._exit(1)
 
     _MODEL_CACHE[model_id] = (model, processor)
     _log(f"model {model_id} loaded")
@@ -59,8 +108,9 @@ 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 +131,8 @@ 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 +205,25 @@ 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]
@@ -48,15 +83,26 @@ def _load_model() -> Any:
     from TTS.api import TTS
 
     _log(f"loading model {MODEL_NAME}")
-    tts = TTS(MODEL_NAME).to(_device())
+    try:
+        tts = TTS(MODEL_NAME).to(_device())
+    except Exception as e:
+        # A load interrupted partway (e.g. OOM) can leave CUDA memory
+        # fragmented/leaked in ways gc.collect()+empty_cache() don't reliably
+        # reclaim, and since _MODEL_CACHE never got populated the idle-unload
+        # loop has nothing to clean up either. Restarting the whole process
+        # is the only guaranteed way to get that memory back — supervisord's
+        # autorestart=true respawns it immediately.
+        _log(f"model load failed, restarting process to reclaim GPU memory: {type(e).__name__}: {e}")
+        os._exit(1)
     _MODEL_CACHE[MODEL_NAME] = tts
     _log("model loaded")
     return tts
 
 
 @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 +120,14 @@ 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 +155,9 @@ 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__":

app/ui/assets/js/asr.js

@@ -7,16 +7,13 @@ async function loadConfig() {
     const r = await fetch('/asr/config');
     SERVER_CONFIG = await r.json();
   } catch (e) {
-    SERVER_CONFIG = {
-      default_backend: 'faster-whisper',
-      default_model: 'large-v3',
-      default_language: 'ko',
-      backends: {
-        'faster-whisper': { models: ['tiny','base','small','medium','large-v3','large-v2','turbo'] },
-        'qwen3': { models: ['Qwen/Qwen3-ASR-0.6B-hf','Qwen/Qwen3-ASR-1.7B-hf'] },
-        'vibevoice': { models: ['microsoft/VibeVoice-ASR'] },
-      },
-    };
+    // No hardcoded fallback list here on purpose: the server's /asr/config is
+    // the single source of truth for backend/model names. A duplicated guess
+    // here would silently drift out of sync with it (this has already bitten
+    // us once — see qwen3 model IDs). Surface the failure instead.
+    SERVER_CONFIG = null;
+    const el = document.getElementById('file-status');
+    if (el) el.textContent = '설정을 불러오지 못했습니다. 페이지를 새로고침해주세요.';
   }
   initAsrUI();
 }

envs/faster_whisper.txt

@@ -2,5 +2,5 @@ fastapi==0.115.0
 uvicorn[standard]==0.30.6
 python-multipart==0.0.9
 faster-whisper==1.1.1
-pyannote.audio>=3.1
+pyannote.audio==4.0.7
 numpy

envs/qwen3.txt

@@ -2,7 +2,7 @@
 fastapi==0.115.0
 uvicorn[standard]==0.30.6
 python-multipart==0.0.9
-transformers>=4.45.0
+transformers==5.14.1
 accelerate>=0.30.0
 librosa>=0.10.0
 soundfile>=0.12.0

scripts/smoke_test.sh

@@ -0,0 +1,50 @@
+#!/usr/bin/env bash
+# Minimal post-deploy smoke test. Hits every ASR/TTS worker's own endpoint
+# directly (bypassing the gateway's OIDC auth) with a synthetic test clip, to
+# catch startup/dependency/API-compat regressions before a user has to find
+# them (qwen3's wrong model IDs, pyannote's use_auth_token/token rename, and
+# the cuDNN LD_LIBRARY_PATH conflict all would have shown up here as an
+# immediate FAIL instead of a silent break discovered later).
+#
+# Run this after every `bash build.sh` + redeploy.
+set -uo pipefail
+
+CONTAINER="${SPEECH_CONTAINER:-speech}"
+TMP_WAV="/tmp/smoke_test_tone.wav"
+FAIL=0
+
+pass() { printf '  OK  %s\n' "$1"; }
+fail() { printf 'FAIL  %s: %s\n' "$1" "$2"; FAIL=1; }
+
+echo "=== generating synthetic test clip ==="
+podman exec "$CONTAINER" ffmpeg -y -f lavfi -i "sine=frequency=440:duration=3" -ar 16000 -ac 1 "$TMP_WAV" >/dev/null 2>&1
+
+echo "=== faster-whisper (8001) ==="
+resp=$(podman exec "$CONTAINER" curl -s -o /dev/null -w "%{http_code}" -X POST http://127.0.0.1:8001/transcribe \
+  -F "file=@${TMP_WAV};type=audio/wav" -F "model=tiny" -F "language=ko")
+[ "$resp" = "200" ] && pass "faster-whisper" || fail "faster-whisper" "HTTP $resp"
+
+echo "=== qwen3 (8004) ==="
+resp=$(podman exec "$CONTAINER" curl -s -o /dev/null -w "%{http_code}" -X POST http://127.0.0.1:8004/transcribe \
+  -F "file=@${TMP_WAV};type=audio/wav" -F "model=Qwen/Qwen3-ASR-0.6B-hf" -F "language=ko")
+[ "$resp" = "200" ] && pass "qwen3" || fail "qwen3" "HTTP $resp"
+
+echo "=== vibevoice (8006) ==="
+resp=$(podman exec "$CONTAINER" curl -s -o /dev/null -w "%{http_code}" -X POST http://127.0.0.1:8006/transcribe \
+  -F "file=@${TMP_WAV};type=audio/wav" -F "model=microsoft/VibeVoice-ASR")
+[ "$resp" = "200" ] && pass "vibevoice" || fail "vibevoice" "HTTP $resp"
+
+echo "=== xtts (8005) ==="
+resp=$(podman exec "$CONTAINER" curl -s -o /dev/null -w "%{http_code}" -X POST http://127.0.0.1:8005/synthesize \
+  -F "text=스모크 테스트입니다." -F "language=ko" -F "speaker_wav=${TMP_WAV}")
+[ "$resp" = "200" ] && pass "xtts" || fail "xtts" "HTTP $resp"
+
+podman exec "$CONTAINER" rm -f "$TMP_WAV"
+
+echo
+if [ "$FAIL" = "0" ]; then
+  echo "all backends OK"
+else
+  echo "one or more backends FAILED — check: podman logs $CONTAINER"
+fi
+exit "$FAIL"