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_whisper.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="WhisperDeployment", num_replicas=1)
 class WhisperDeployment:
@@ -38,6 +41,17 @@ class WhisperDeployment:
 
         print("Whisper model loaded successfully")
 
+        # MLflow metrics
+        self._mlflow = InferenceLogger(
+            experiment_name="ray-serve-whisper",
+            run_name=f"whisper-{self.model_size}",
+            tags={"model.name": f"whisper-{self.model_size}", "model.framework": "faster-whisper", "device": self.device},
+            flush_every=5,
+        )
+        self._mlflow.initialize(
+            params={"model_size": self.model_size, "device": self.device, "compute_type": self.compute_type}
+        )
+
     async def __call__(self, request: dict[str, Any]) -> dict[str, Any]:
         """
         Handle transcription requests.
@@ -59,6 +73,7 @@ class WhisperDeployment:
         }
         """
 
+        _start = time.time()
         language = request.get("language")
         task = request.get("task", "transcribe")  # transcribe or translate
         response_format = request.get("response_format", "json")
@@ -130,6 +145,14 @@ class WhisperDeployment:
                 "segments": segment_list,
             }
 
+        # Log to MLflow
+        self._mlflow.log_request(
+            latency_s=time.time() - _start,
+            audio_duration_s=info.duration,
+            segments=len(segment_list),
+            realtime_factor=(time.time() - _start) / info.duration if info.duration > 0 else 0,
+        )
+
         # Default JSON format (OpenAI-compatible)
         return {
             "text": full_text.strip(),