fix: create the metrics tracker only on trainer rank zero

specforge/training/assembly.py

@@ -292,7 +292,15 @@ def _logger(metrics, step):
 
 
 def _configured_logger(cfg: Config):
-    """Create an external tracker only for a trainer-bearing run."""
+    """Create an external tracker only on the trainer's global rank zero."""
+    if cfg.training.role != "producer":
+        import torch.distributed as dist
+
+        # A distributed run has one logical metric stream. Letting every rank
+        # create W&B/MLflow runs or write the same TensorBoard directory both
+        # duplicates metrics and makes large jobs increasingly fragile.
+        if dist.is_available() and dist.is_initialized() and dist.get_rank() != 0:
+            return None
     if cfg.tracking.report_to == "none" or cfg.training.role == "producer":
         return _logger
 

tests/test_config/test_unified_feature_reachability.py

@@ -205,6 +205,15 @@ def test_tracking_config_reaches_the_existing_tracker_adapter(self):
         self.assertEqual(output_dir, "/tmp/output")
         self.assertIs(create.call_args.kwargs["console_logger"], _logger)
 
+    def test_only_global_rank_zero_creates_a_training_logger(self):
+        cfg = Config.model_validate(OFFLINE_EAGLE3)
+        with (
+            mock.patch("torch.distributed.is_available", return_value=True),
+            mock.patch("torch.distributed.is_initialized", return_value=True),
+            mock.patch("torch.distributed.get_rank", return_value=3),
+        ):
+            self.assertIsNone(_configured_logger(cfg))
+
     def test_tracking_backend_is_strictly_typed(self):
         with self.assertRaises(ValidationError):
             Config.model_validate(