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

@@ -167,12 +167,23 @@ def _load_model(model_name: str) -> WhisperModel:
         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(
+        try:
-            model_name,
+            _MODEL_CACHE[model_name] = WhisperModel(
-            device=DEVICE,
+                model_name,
-            compute_type=COMPUTE_TYPE,
+                device=DEVICE,
-            download_root=str(MODEL_CACHE),
+                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

@@ -109,15 +109,25 @@ def _load_model(model_id: str) -> Tuple[Any, Any]:
     _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))
+    try:
-    model = AutoModelForSpeechSeq2Seq.from_pretrained(
+        processor = AutoProcessor.from_pretrained(model_id, cache_dir=str(MODEL_CACHE))
-        model_id,
+        model = AutoModelForSpeechSeq2Seq.from_pretrained(
-        dtype=dtype,
+            model_id,
-        low_cpu_mem_usage=True,
+            dtype=dtype,
-        cache_dir=str(MODEL_CACHE),
+            low_cpu_mem_usage=True,
-    )
+            cache_dir=str(MODEL_CACHE),
-    model.to(DEVICE)
+        )
-    model.eval()
+        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")

app/asr/workers/vibevoice_worker.py

@@ -77,20 +77,30 @@ def _load_model(model_id: str) -> Any:
     from vibevoice.processor.vibevoice_asr_processor import VibeVoiceASRProcessor
 
     _log(f"loading model {model_id}")
-    processor = VibeVoiceASRProcessor.from_pretrained(
+    try:
-        model_id,
+        processor = VibeVoiceASRProcessor.from_pretrained(
-        language_model_pretrained_name="Qwen/Qwen2.5-1.5B",
+            model_id,
-        cache_dir=str(MODEL_CACHE),
+            language_model_pretrained_name="Qwen/Qwen2.5-1.5B",
-    )
+            cache_dir=str(MODEL_CACHE),
-    model = VibeVoiceASRForConditionalGeneration.from_pretrained(
+        )
-        model_id,
+        model = VibeVoiceASRForConditionalGeneration.from_pretrained(
-        dtype=torch.bfloat16 if DEVICE == "cuda" else torch.float32,
+            model_id,
-        attn_implementation="sdpa",
+            dtype=torch.bfloat16 if DEVICE == "cuda" else torch.float32,
-        trust_remote_code=True,
+            attn_implementation="sdpa",
-        cache_dir=str(MODEL_CACHE),
+            trust_remote_code=True,
-    )
+            cache_dir=str(MODEL_CACHE),
-    model.to(DEVICE)
+        )
-    model.eval()
+        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")

app/tts/workers/xtts_worker.py

@@ -83,7 +83,17 @@ 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