feat: add MLflow experiment tracking to all 4 Gradio UIs

Each UI now logs per-request metrics to MLflow:
- llm.py: latency, tokens/sec, prompt/completion tokens (gradio-llm-tuning)
- embeddings.py: latency, text length, batch size (gradio-embeddings-tuning)
- stt.py: latency, audio duration, real-time factor (gradio-stt-tuning)
- tts.py: latency, text length, audio duration (gradio-tts-tuning)

Uses try/except guarded imports so UIs still work if MLflow is
unreachable. Persistent run per Gradio instance, batched metric logging
via MlflowClient.log_batch().

tts.py

@@ -37,6 +37,66 @@ MLFLOW_TRACKING_URI = os.environ.get(
     "http://mlflow.mlflow.svc.cluster.local:80"
 )
 
+# ─── MLflow experiment tracking ──────────────────────────────────────────
+try:
+    import mlflow
+    from mlflow.tracking import MlflowClient
+
+    mlflow.set_tracking_uri(MLFLOW_TRACKING_URI)
+    _mlflow_client = MlflowClient()
+
+    _experiment = _mlflow_client.get_experiment_by_name("gradio-tts-tuning")
+    if _experiment is None:
+        _experiment_id = _mlflow_client.create_experiment(
+            "gradio-tts-tuning",
+            artifact_location="/mlflow/artifacts/gradio-tts-tuning",
+        )
+    else:
+        _experiment_id = _experiment.experiment_id
+
+    _mlflow_run = mlflow.start_run(
+        experiment_id=_experiment_id,
+        run_name=f"gradio-tts-{os.environ.get('HOSTNAME', 'local')}",
+        tags={"service": "gradio-tts", "endpoint": TTS_URL},
+    )
+    _mlflow_run_id = _mlflow_run.info.run_id
+    _mlflow_step = 0
+    MLFLOW_ENABLED = True
+    logger.info("MLflow tracking enabled: experiment=%s run=%s", _experiment_id, _mlflow_run_id)
+except Exception as exc:
+    logger.warning("MLflow tracking disabled: %s", exc)
+    _mlflow_client = None
+    _mlflow_run_id = None
+    _mlflow_step = 0
+    MLFLOW_ENABLED = False
+
+
+def _log_tts_metrics(
+    latency: float, audio_duration: float, text_chars: int, language: str,
+) -> None:
+    """Log TTS inference metrics to MLflow (non-blocking best-effort)."""
+    global _mlflow_step
+    if not MLFLOW_ENABLED or _mlflow_client is None:
+        return
+    try:
+        _mlflow_step += 1
+        ts = int(time.time() * 1000)
+        rtf = latency / audio_duration if audio_duration > 0 else 0
+        cps = text_chars / latency if latency > 0 else 0
+        _mlflow_client.log_batch(
+            _mlflow_run_id,
+            metrics=[
+                mlflow.entities.Metric("latency_s", latency, ts, _mlflow_step),
+                mlflow.entities.Metric("audio_duration_s", audio_duration, ts, _mlflow_step),
+                mlflow.entities.Metric("realtime_factor", rtf, ts, _mlflow_step),
+                mlflow.entities.Metric("chars_per_second", cps, ts, _mlflow_step),
+                mlflow.entities.Metric("text_chars", text_chars, ts, _mlflow_step),
+            ],
+        )
+    except Exception:
+        logger.debug("MLflow log failed", exc_info=True)
+
+
 # HTTP client with longer timeout for audio generation
 client = httpx.Client(timeout=120.0)
 
@@ -93,6 +153,14 @@ def synthesize_speech(text: str, language: str) -> tuple[str, tuple[int, np.ndar
         
         # Status message
         status = f"✅ Generated {duration:.2f}s of audio in {latency*1000:.0f}ms"
+
+        # Log to MLflow
+        _log_tts_metrics(
+            latency=latency,
+            audio_duration=duration,
+            text_chars=len(text),
+            language=lang_code,
+        )
         
         # Metrics
         metrics = f"""