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_embeddings.py

@@ -4,10 +4,13 @@ Runs on: drizzt (Radeon 680M iGPU, ROCm)
 """
 
 import os
+import time
 from typing import Any
 
 from ray import serve
 
+from ray_serve.mlflow_logger import InferenceLogger
+
 
 @serve.deployment(name="EmbeddingsDeployment", num_replicas=1)
 class EmbeddingsDeployment:
@@ -33,6 +36,17 @@ class EmbeddingsDeployment:
 
         print(f"Model loaded. Embedding dimension: {self.embedding_dim}")
 
+        # MLflow metrics
+        self._mlflow = InferenceLogger(
+            experiment_name="ray-serve-embeddings",
+            run_name=f"embeddings-{self.model_id.split('/')[-1]}",
+            tags={"model.name": self.model_id, "model.framework": "sentence-transformers", "device": self.device},
+            flush_every=10,
+        )
+        self._mlflow.initialize(
+            params={"model_id": self.model_id, "embedding_dim": str(self.embedding_dim), "device": self.device}
+        )
+
     async def __call__(self, request: dict[str, Any]) -> dict[str, Any]:
         """
         Handle OpenAI-compatible embedding requests.
@@ -46,6 +60,8 @@ class EmbeddingsDeployment:
         """
         input_data = request.get("input", "")
 
+        _start = time.time()
+
         # Handle both single string and list of strings
         texts = [input_data] if isinstance(input_data, str) else input_data
 
@@ -69,6 +85,13 @@ class EmbeddingsDeployment:
             )
             total_tokens += len(text.split())
 
+        # Log to MLflow
+        self._mlflow.log_request(
+            latency_s=time.time() - _start,
+            batch_size=len(texts),
+            total_tokens=total_tokens,
+        )
+
         # Return OpenAI-compatible response
         return {
             "object": "list",