Fix MLflowTracker.store_init_configuration mutating the caller's config dict

src/accelerate/tracking.py

@@ -795,16 +795,16 @@ def store_init_configuration(self, values: dict):
         """
         import mlflow
 
-        for name, value in list(values.items()):
+        values_list = []
+        for name, value in values.items():
             # internally, all values are converted to str in MLflow
             if len(str(value)) > mlflow.utils.validation.MAX_PARAM_VAL_LENGTH:
                 logger.warning_once(
                     f'Accelerate is attempting to log a value of "{value}" for key "{name}" as a parameter. MLflow\'s'
                     f" log_param() only accepts values no longer than {mlflow.utils.validation.MAX_PARAM_VAL_LENGTH} characters so we dropped this attribute."
                 )
-                del values[name]
-
-        values_list = list(values.items())
+            else:
+                values_list.append((name, value))
 
         # MLflow cannot log more than 100 values in one go, so we have to split it
         for i in range(0, len(values_list), mlflow.utils.validation.MAX_PARAMS_TAGS_PER_BATCH):

tests/test_tracking.py

@@ -322,6 +322,32 @@ def test_log_artifacts(self):
             ],
         )
 
+    def test_store_init_configuration_does_not_mutate_input(self):
+        import mlflow
+
+        """`store_init_configuration` must not mutate the caller's dict when dropping over-long values.
+
+        The same `config` object is shared across all trackers in `Accelerator.init_trackers`, so deleting
+        keys in place removed them from the user's dict and from every other tracker too.
+        """
+        too_long = "x" * (mlflow.utils.validation.MAX_PARAM_VAL_LENGTH + 1)
+        values = {"learning_rate": 0.001, "too_long": too_long}
+        tracker = MLflowTracker(experiment_name="test_exp", logging_dir=self.tmpdir.name)
+        accelerator = Accelerator(log_with=tracker)
+        accelerator.init_trackers(project_name="test_exp")
+        tracker.store_init_configuration(values)
+
+        run_id = tracker.active_run.info.run_id
+        accelerator.end_training()
+
+        # The caller's dict must be left untouched (previously `too_long` was deleted in place).
+        self.assertEqual(values, {"learning_rate": 0.001, "too_long": too_long})
+
+        # The over-long value is still dropped from what is actually logged to MLflow.
+        params = mlflow.get_run(run_id).data.params
+        self.assertIn("learning_rate", params)
+        self.assertNotIn("too_long", params)
+
 
 @require_comet_ml
 class CometMLTest(unittest.TestCase):