⚡ Bolt: Add Parakeet model warmup to reduce first-inference latency

.jules/bolt.md

@@ -1,3 +1,7 @@
 ## 2024-02-18 - Pre-scaled Audio Feedback
 **Learning:** AudioFeedback volume scaling was applied during every playback, causing unnecessary numpy overhead and latency.
 **Action:** Pre-calculate scaled audio during `_load_and_cache` to minimize `_play_cached` latency (from ~1ms to ~0.04ms).
+
+## 2024-05-22 - Parakeet Model Cold Start
+**Learning:** ONNX Runtime initialization and graph optimization cause significant latency (cold start) on the first inference.
+**Action:** Implemented `ParakeetManager.warmup()` to run a dummy inference at startup, shifting latency away from the first user interaction.

src/chirp/main.py

@@ -72,6 +72,7 @@ def __init__(self, *, verbose: bool = False) -> None:
                     model_dir=model_dir,
                     timeout=self.config.model_timeout,
                 )
+                self.parakeet.warmup()
         except ModelNotPreparedError as exc:
             self.logger.error(str(exc))
             raise SystemExit(1) from exc

src/chirp/parakeet_manager.py

@@ -123,6 +123,15 @@ def _load_model(self):
                 f"Model not found at {self._model_dir} — run: uv run chirp-setup"
             ) from exc
 
+    def warmup(self) -> None:
+        """Run a dummy inference to initialize internal buffers and optimize execution graph."""
+        self._logger.debug("Warming up Parakeet model...")
+        dummy_audio = np.zeros(16_000, dtype=np.float32)
+        try:
+            self.transcribe(dummy_audio)
+        except Exception as exc:
+            self._logger.warning("Warmup failed (non-fatal): %s", exc)
+
     def transcribe(self, audio: np.ndarray, *, sample_rate: int = 16_000, language: Optional[str] = None) -> str:
         with self._lock:
             self._last_access = time.time()

tests/test_parakeet_manager.py

@@ -107,6 +107,29 @@ def test_timeout_zero_disables_monitor(self, mock_onnx):
         self.assertIsNone(manager._monitor_thread)
         self.assertIsNotNone(manager._model)
 
+    @patch("chirp.parakeet_manager.onnx_asr")
+    def test_warmup(self, mock_onnx):
+        """Test that warmup calls transcribe with dummy audio."""
+        mock_onnx.load_model.return_value = MagicMock()
+        manager = ParakeetManager(
+            model_name="test",
+            quantization=None,
+            provider_key="cpu",
+            threads=1,
+            logger=self.logger,
+            model_dir=self.model_dir,
+            timeout=100.0,
+        )
+
+        # We can mock transcribe to ensure it's called
+        with patch.object(manager, 'transcribe') as mock_transcribe:
+            manager.warmup()
+            mock_transcribe.assert_called_once()
+            args, _ = mock_transcribe.call_args
+            audio_arg = args[0]
+            self.assertEqual(audio_arg.shape, (16000,))
+            self.assertEqual(audio_arg.dtype, np.float32)
+
 
 if __name__ == "__main__":
     unittest.main()