feat: add MLflow inference logging to all Ray Serve apps

- Add mlflow_logger.py: lightweight REST-based MLflow logger (no mlflow dep)
- Instrument serve_llm.py with latency, token counts, tokens/sec metrics
- Instrument serve_embeddings.py with latency, batch_size, total_tokens
- Instrument serve_whisper.py with latency, audio_duration, realtime_factor
- Instrument serve_tts.py with latency, audio_duration, text_chars
- Instrument serve_reranker.py with latency, num_pairs, top_k

ray_serve/serve_tts.py

@@ -6,10 +6,13 @@ Runs on: elminster (RTX 2070 8GB, CUDA)
 import base64
 import io
 import os
+import time
 from typing import Any
 
 from ray import serve
 
+from ray_serve.mlflow_logger import InferenceLogger
+
 
 @serve.deployment(name="TTSDeployment", num_replicas=1)
 class TTSDeployment:
@@ -32,6 +35,15 @@ class TTSDeployment:
 
         print("TTS model loaded successfully")
 
+        # MLflow metrics
+        self._mlflow = InferenceLogger(
+            experiment_name="ray-serve-tts",
+            run_name=f"tts-{self.model_name.split('/')[-1]}",
+            tags={"model.name": self.model_name, "model.framework": "coqui-tts", "gpu": str(self.use_gpu)},
+            flush_every=5,
+        )
+        self._mlflow.initialize(params={"model_name": self.model_name, "use_gpu": str(self.use_gpu)})
+
     async def __call__(self, request: dict[str, Any]) -> dict[str, Any]:
         """
         Handle text-to-speech requests.
@@ -49,6 +61,7 @@ class TTSDeployment:
         import numpy as np
         from scipy.io import wavfile
 
+        _start = time.time()
         text = request.get("text", "")
         speaker = request.get("speaker")
         language = request.get("language")
@@ -88,10 +101,20 @@ class TTSDeployment:
             wavfile.write(buffer, sample_rate, wav_int16)
             audio_bytes = buffer.getvalue()
 
+            duration = len(wav) / sample_rate
+
+            # Log to MLflow
+            self._mlflow.log_request(
+                latency_s=time.time() - _start,
+                audio_duration_s=duration,
+                text_chars=len(text),
+                realtime_factor=(time.time() - _start) / duration if duration > 0 else 0,
+            )
+
             response = {
                 "model": self.model_name,
                 "sample_rate": sample_rate,
-                "duration": len(wav) / sample_rate,
+                "duration": duration,
                 "format": output_format,
             }