diff --git a/examples/pytorch/language-modeling/run_clm_no_trainer.py b/examples/pytorch/language-modeling/run_clm_no_trainer.py
index 3435ea2ecdbd..c7976ef350e1 100755
--- a/examples/pytorch/language-modeling/run_clm_no_trainer.py
+++ b/examples/pytorch/language-modeling/run_clm_no_trainer.py
@@ -200,7 +200,7 @@ def parse_args():
     )
     parser.add_argument(
         "--with_tracking",
-        required=False,
+        action="store_true",
         help="Whether to load in all available experiment trackers from the environment and use them for logging.",
     )
     args = parser.parse_args()
@@ -227,7 +227,7 @@ def main():
 
     # Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
     # If we're using tracking, we also need to initialize it here and it will pick up all supported trackers in the environment
-    accelerator = Accelerator(log_with="all") if args.with_tracking else Accelerator()
+    accelerator = Accelerator(log_with="all", logging_dir=args.output_dir) if args.with_tracking else Accelerator()
     # Make one log on every process with the configuration for debugging.
     logging.basicConfig(
         format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
@@ -485,7 +485,10 @@ def group_texts(examples):
 
     # We need to initialize the trackers we use, and also store our configuration
     if args.with_tracking:
-        accelerator.init_trackers("clm_no_trainer", args)
+        experiment_config = vars(args)
+        # TensorBoard cannot log Enums, need the raw value
+        experiment_config["lr_scheduler_type"] = experiment_config["lr_scheduler_type"].value
+        accelerator.init_trackers("clm_no_trainer", experiment_config)
 
     # Train!
     total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
@@ -571,12 +574,7 @@ def group_texts(examples):
 
         if args.with_tracking:
             accelerator.log(
-                {
-                    "perplexity": perplexity,
-                    "train_loss": total_loss,
-                    "epoch": epoch,
-                },
-                step=completed_steps,
+                {"perplexity": perplexity, "train_loss": total_loss, "epoch": epoch, "step": completed_steps},
             )
 
         if args.push_to_hub and epoch < args.num_train_epochs - 1:
diff --git a/examples/pytorch/language-modeling/run_mlm_no_trainer.py b/examples/pytorch/language-modeling/run_mlm_no_trainer.py
index 2720e76d02e9..6ff58ddc7f32 100755
--- a/examples/pytorch/language-modeling/run_mlm_no_trainer.py
+++ b/examples/pytorch/language-modeling/run_mlm_no_trainer.py
@@ -209,7 +209,7 @@ def parse_args():
     )
     parser.add_argument(
         "--with_tracking",
-        required=False,
+        action="store_true",
         help="Whether to load in all available experiment trackers from the environment and use them for logging.",
     )
     args = parser.parse_args()
@@ -238,7 +238,7 @@ def main():
 
     # Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
     # If we're using tracking, we also need to initialize it here and it will pick up all supported trackers in the environment
-    accelerator = Accelerator(log_with="all") if args.with_tracking else Accelerator()
+    accelerator = Accelerator(log_with="all", logging_dir=args.output_dir) if args.with_tracking else Accelerator()
     # Make one log on every process with the configuration for debugging.
     logging.basicConfig(
         format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
@@ -531,7 +531,10 @@ def group_texts(examples):
 
     # We need to initialize the trackers we use, and also store our configuration
     if args.with_tracking:
-        accelerator.init_trackers("clm_no_trainer", args)
+        experiment_config = vars(args)
+        # TensorBoard cannot log Enums, need the raw value
+        experiment_config["lr_scheduler_type"] = experiment_config["lr_scheduler_type"].value
+        accelerator.init_trackers("mlm_no_trainer", experiment_config)
 
     # Train!
     total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
@@ -618,12 +621,7 @@ def group_texts(examples):
 
         if args.with_tracking:
             accelerator.log(
-                {
-                    "perplexity": perplexity,
-                    "train_loss": total_loss,
-                    "epoch": epoch,
-                },
-                step=completed_steps,
+                {"perplexity": perplexity, "train_loss": total_loss, "epoch": epoch, "step": completed_steps},
             )
 
         if args.push_to_hub and epoch < args.num_train_epochs - 1:
diff --git a/examples/pytorch/multiple-choice/run_swag_no_trainer.py b/examples/pytorch/multiple-choice/run_swag_no_trainer.py
index f845cd43e2c8..63b799c0929b 100755
--- a/examples/pytorch/multiple-choice/run_swag_no_trainer.py
+++ b/examples/pytorch/multiple-choice/run_swag_no_trainer.py
@@ -192,7 +192,7 @@ def parse_args():
     )
     parser.add_argument(
         "--with_tracking",
-        required=False,
+        action="store_true",
         help="Whether to load in all available experiment trackers from the environment and use them for logging.",
     )
     args = parser.parse_args()
@@ -265,7 +265,7 @@ def main():
 
     # Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
     # If we're using tracking, we also need to initialize it here and it will pick up all supported trackers in the environment
-    accelerator = Accelerator(log_with="all") if args.with_tracking else Accelerator()
+    accelerator = Accelerator(log_with="all", logging_dir=args.output_dir) if args.with_tracking else Accelerator()
     # Make one log on every process with the configuration for debugging.
     logging.basicConfig(
         format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
@@ -485,7 +485,10 @@ def preprocess_function(examples):
 
     # We need to initialize the trackers we use, and also store our configuration
     if args.with_tracking:
-        accelerator.init_trackers("clm_no_trainer", args)
+        experiment_config = vars(args)
+        # TensorBoard cannot log Enums, need the raw value
+        experiment_config["lr_scheduler_type"] = experiment_config["lr_scheduler_type"].value
+        accelerator.init_trackers("swag_no_trainer", experiment_config)
 
     # Metrics
     metric = load_metric("accuracy")
@@ -570,12 +573,7 @@ def preprocess_function(examples):
 
         if args.with_tracking:
             accelerator.log(
-                {
-                    "accuracy": eval_metric,
-                    "train_loss": total_loss,
-                    "epoch": epoch,
-                },
-                step=completed_steps,
+                {"accuracy": eval_metric, "train_loss": total_loss, "epoch": epoch, "step": completed_steps},
             )
 
         if args.push_to_hub and epoch < args.num_train_epochs - 1:
diff --git a/examples/pytorch/question-answering/run_qa_beam_search_no_trainer.py b/examples/pytorch/question-answering/run_qa_beam_search_no_trainer.py
index bdfa1e189322..a53a3cf54822 100644
--- a/examples/pytorch/question-answering/run_qa_beam_search_no_trainer.py
+++ b/examples/pytorch/question-answering/run_qa_beam_search_no_trainer.py
@@ -224,7 +224,7 @@ def parse_args():
     )
     parser.add_argument(
         "--with_tracking",
-        required=False,
+        action="store_true",
         help="Whether to load in all available experiment trackers from the environment and use them for logging.",
     )
     args = parser.parse_args()
@@ -259,7 +259,7 @@ def main():
 
     # Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
     # If we're using tracking, we also need to initialize it here and it will pick up all supported trackers in the environment
-    accelerator = Accelerator(log_with="all") if args.with_tracking else Accelerator()
+    accelerator = Accelerator(log_with="all", logging_dir=args.output_dir) if args.with_tracking else Accelerator()
     # Make one log on every process with the configuration for debugging.
     logging.basicConfig(
         format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
@@ -723,7 +723,10 @@ def create_and_fill_np_array(start_or_end_logits, dataset, max_len):
 
     # We need to initialize the trackers we use, and also store our configuration
     if args.with_tracking:
-        accelerator.init_trackers("clm_no_trainer", args)
+        experiment_config = vars(args)
+        # TensorBoard cannot log Enums, need the raw value
+        experiment_config["lr_scheduler_type"] = experiment_config["lr_scheduler_type"].value
+        accelerator.init_trackers("qa_beam_search_no_trainer", experiment_config)
 
     # Train!
     total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
@@ -916,11 +919,12 @@ def create_and_fill_np_array(start_or_end_logits, dataset, max_len):
             "squad_v2" if args.version_2_with_negative else "squad": eval_metric,
             "train_loss": total_loss,
             "epoch": epoch,
+            "step": completed_steps,
         }
         if args.do_predict:
             log["squad_v2_predict" if args.version_2_with_negative else "squad_predict"] = predict_metric
 
-        accelerator.log(log, step=completed_steps)
+        accelerator.log(log)
 
     if args.checkpointing_steps == "epoch":
         accelerator.save_state(f"epoch_{epoch}")
diff --git a/examples/pytorch/question-answering/run_qa_no_trainer.py b/examples/pytorch/question-answering/run_qa_no_trainer.py
index 925c31b70642..7c7b54004ff5 100755
--- a/examples/pytorch/question-answering/run_qa_no_trainer.py
+++ b/examples/pytorch/question-answering/run_qa_no_trainer.py
@@ -254,7 +254,7 @@ def parse_args():
     )
     parser.add_argument(
         "--with_tracking",
-        required=False,
+        action="store_true",
         help="Whether to load in all available experiment trackers from the environment and use them for logging.",
     )
     args = parser.parse_args()
@@ -289,7 +289,7 @@ def main():
 
     # Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
     # If we're using tracking, we also need to initialize it here and it will pick up all supported trackers in the environment
-    accelerator = Accelerator(log_with="all") if args.with_tracking else Accelerator()
+    accelerator = Accelerator(log_with="all", logging_dir=args.output_dir) if args.with_tracking else Accelerator()
     # Make one log on every process with the configuration for debugging.
     logging.basicConfig(
         format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
@@ -730,7 +730,10 @@ def create_and_fill_np_array(start_or_end_logits, dataset, max_len):
 
     # We need to initialize the trackers we use, and also store our configuration
     if args.with_tracking:
-        accelerator.init_trackers("clm_no_trainer", args)
+        experiment_config = vars(args)
+        # TensorBoard cannot log Enums, need the raw value
+        experiment_config["lr_scheduler_type"] = experiment_config["lr_scheduler_type"].value
+        accelerator.init_trackers("qa_no_trainer", experiment_config)
 
     # Train!
     total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
@@ -889,11 +892,12 @@ def create_and_fill_np_array(start_or_end_logits, dataset, max_len):
             "squad_v2" if args.version_2_with_negative else "squad": eval_metric,
             "train_loss": total_loss,
             "epoch": epoch,
+            "step": completed_steps,
         }
     if args.do_predict:
         log["squad_v2_predict" if args.version_2_with_negative else "squad_predict"] = predict_metric
 
-        accelerator.log(log, step=completed_steps)
+        accelerator.log(log)
 
     if args.output_dir is not None:
         accelerator.wait_for_everyone()
diff --git a/examples/pytorch/summarization/run_summarization_no_trainer.py b/examples/pytorch/summarization/run_summarization_no_trainer.py
index 1c93b064a6d1..48c0aada34b4 100644
--- a/examples/pytorch/summarization/run_summarization_no_trainer.py
+++ b/examples/pytorch/summarization/run_summarization_no_trainer.py
@@ -277,7 +277,7 @@ def parse_args():
     )
     parser.add_argument(
         "--with_tracking",
-        required=False,
+        action="store_true",
         help="Whether to load in all available experiment trackers from the environment and use them for logging.",
     )
     args = parser.parse_args()
@@ -315,7 +315,7 @@ def main():
         )
     # Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
     # If we're using tracking, we also need to initialize it here and it will pick up all supported trackers in the environment
-    accelerator = Accelerator(log_with="all") if args.with_tracking else Accelerator()
+    accelerator = Accelerator(log_with="all", logging_dir=args.output_dir) if args.with_tracking else Accelerator()
     # Make one log on every process with the configuration for debugging.
     logging.basicConfig(
         format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
@@ -548,7 +548,10 @@ def postprocess_text(preds, labels):
 
     # We need to initialize the trackers we use, and also store our configuration
     if args.with_tracking:
-        accelerator.init_trackers("summarization_no_trainer", args)
+        experiment_config = vars(args)
+        # TensorBoard cannot log Enums, need the raw value
+        experiment_config["lr_scheduler_type"] = experiment_config["lr_scheduler_type"].value
+        accelerator.init_trackers("summarization_no_trainer", experiment_config)
 
     # Metric
     metric = load_metric("rouge")
@@ -666,7 +669,8 @@ def postprocess_text(preds, labels):
         if args.with_tracking:
             result["train_loss"] = total_loss
             result["epoch"] = epoch
-            accelerator.log(result, step=completed_steps)
+            result["step"] = completed_steps
+            accelerator.log(result)
 
         if args.push_to_hub and epoch < args.num_train_epochs - 1:
             accelerator.wait_for_everyone()
diff --git a/examples/pytorch/test_accelerate_examples.py b/examples/pytorch/test_accelerate_examples.py
index f4fb8d0d52a6..720a6f707914 100644
--- a/examples/pytorch/test_accelerate_examples.py
+++ b/examples/pytorch/test_accelerate_examples.py
@@ -104,7 +104,8 @@ def test_run_glue_no_trainer(self):
             --learning_rate=1e-4
             --seed=42
             --checkpointing_steps epoch
-            """.split()
+            --with_tracking
+        """.split()
 
         if is_cuda_and_apex_available():
             testargs.append("--fp16")
@@ -114,6 +115,7 @@ def test_run_glue_no_trainer(self):
             result = get_results(tmp_dir)
             self.assertGreaterEqual(result["eval_accuracy"], 0.75)
             self.assertTrue(os.path.exists(os.path.join(tmp_dir, "epoch_0")))
+            self.assertTrue(os.path.exists(os.path.join(tmp_dir, "glue_no_trainer")))
 
     def test_run_clm_no_trainer(self):
         tmp_dir = self.get_auto_remove_tmp_dir()
@@ -128,7 +130,8 @@ def test_run_clm_no_trainer(self):
             --num_train_epochs 2
             --output_dir {tmp_dir}
             --checkpointing_steps epoch
-            """.split()
+            --with_tracking
+        """.split()
 
         if torch.cuda.device_count() > 1:
             # Skipping because there are not enough batches to train the model + would need a drop_last to work.
@@ -139,6 +142,7 @@ def test_run_clm_no_trainer(self):
             result = get_results(tmp_dir)
             self.assertLess(result["perplexity"], 100)
             self.assertTrue(os.path.exists(os.path.join(tmp_dir, "epoch_0")))
+            self.assertTrue(os.path.exists(os.path.join(tmp_dir, "clm_no_trainer")))
 
     def test_run_mlm_no_trainer(self):
         tmp_dir = self.get_auto_remove_tmp_dir()
@@ -150,6 +154,7 @@ def test_run_mlm_no_trainer(self):
             --output_dir {tmp_dir}
             --num_train_epochs=1
             --checkpointing_steps epoch
+            --with_tracking
         """.split()
 
         with patch.object(sys, "argv", testargs):
@@ -157,6 +162,7 @@ def test_run_mlm_no_trainer(self):
             result = get_results(tmp_dir)
             self.assertLess(result["perplexity"], 42)
             self.assertTrue(os.path.exists(os.path.join(tmp_dir, "epoch_0")))
+            self.assertTrue(os.path.exists(os.path.join(tmp_dir, "mlm_no_trainer")))
 
     def test_run_ner_no_trainer(self):
         # with so little data distributed training needs more epochs to get the score on par with 0/1 gpu
@@ -175,6 +181,7 @@ def test_run_ner_no_trainer(self):
             --num_train_epochs={epochs}
             --seed 7
             --checkpointing_steps epoch
+            --with_tracking
         """.split()
 
         with patch.object(sys, "argv", testargs):
@@ -183,6 +190,7 @@ def test_run_ner_no_trainer(self):
             self.assertGreaterEqual(result["eval_accuracy"], 0.75)
             self.assertLess(result["train_loss"], 0.5)
             self.assertTrue(os.path.exists(os.path.join(tmp_dir, "epoch_0")))
+            self.assertTrue(os.path.exists(os.path.join(tmp_dir, "ner_no_trainer")))
 
     def test_run_squad_no_trainer(self):
         tmp_dir = self.get_auto_remove_tmp_dir()
@@ -199,6 +207,7 @@ def test_run_squad_no_trainer(self):
             --per_device_train_batch_size=2
             --per_device_eval_batch_size=1
             --checkpointing_steps epoch
+            --with_tracking
         """.split()
 
         with patch.object(sys, "argv", testargs):
@@ -207,6 +216,7 @@ def test_run_squad_no_trainer(self):
             self.assertGreaterEqual(result["eval_f1"], 30)
             self.assertGreaterEqual(result["eval_exact"], 30)
             self.assertTrue(os.path.exists(os.path.join(tmp_dir, "epoch_0")))
+            self.assertTrue(os.path.exists(os.path.join(tmp_dir, "qa_no_trainer")))
 
     def test_run_swag_no_trainer(self):
         tmp_dir = self.get_auto_remove_tmp_dir()
@@ -221,12 +231,14 @@ def test_run_swag_no_trainer(self):
             --learning_rate=2e-4
             --per_device_train_batch_size=2
             --per_device_eval_batch_size=1
+            --with_tracking
         """.split()
 
         with patch.object(sys, "argv", testargs):
             run_swag_no_trainer.main()
             result = get_results(tmp_dir)
             self.assertGreaterEqual(result["eval_accuracy"], 0.8)
+            self.assertTrue(os.path.exists(os.path.join(tmp_dir, "swag_no_trainer")))
 
     @slow
     def test_run_summarization_no_trainer(self):
@@ -243,6 +255,7 @@ def test_run_summarization_no_trainer(self):
             --per_device_train_batch_size=2
             --per_device_eval_batch_size=1
             --checkpointing_steps epoch
+            --with_tracking
         """.split()
 
         with patch.object(sys, "argv", testargs):
@@ -253,6 +266,7 @@ def test_run_summarization_no_trainer(self):
             self.assertGreaterEqual(result["eval_rougeL"], 7)
             self.assertGreaterEqual(result["eval_rougeLsum"], 7)
             self.assertTrue(os.path.exists(os.path.join(tmp_dir, "epoch_0")))
+            self.assertTrue(os.path.exists(os.path.join(tmp_dir, "summarization_no_trainer")))
 
     @slow
     def test_run_translation_no_trainer(self):
@@ -273,6 +287,7 @@ def test_run_translation_no_trainer(self):
             --source_lang en_XX
             --target_lang ro_RO
             --checkpointing_steps epoch
+            --with_tracking
         """.split()
 
         with patch.object(sys, "argv", testargs):
@@ -280,3 +295,4 @@ def test_run_translation_no_trainer(self):
             result = get_results(tmp_dir)
             self.assertGreaterEqual(result["eval_bleu"], 30)
             self.assertTrue(os.path.exists(os.path.join(tmp_dir, "epoch_0")))
+            self.assertTrue(os.path.exists(os.path.join(tmp_dir, "translation_no_trainer")))
diff --git a/examples/pytorch/text-classification/run_glue_no_trainer.py b/examples/pytorch/text-classification/run_glue_no_trainer.py
index 9730c2f34568..0842b462f9cb 100644
--- a/examples/pytorch/text-classification/run_glue_no_trainer.py
+++ b/examples/pytorch/text-classification/run_glue_no_trainer.py
@@ -165,7 +165,7 @@ def parse_args():
     )
     parser.add_argument(
         "--with_tracking",
-        required=False,
+        action="store_true",
         help="Whether to load in all available experiment trackers from the environment and use them for logging.",
     )
     args = parser.parse_args()
@@ -192,7 +192,7 @@ def main():
 
     # Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
     # If we're using tracking, we also need to initialize it here and it will pick up all supported trackers in the environment
-    accelerator = Accelerator(log_with="all") if args.with_tracking else Accelerator()
+    accelerator = Accelerator(log_with="all", logging_dir=args.output_dir) if args.with_tracking else Accelerator()
     # Make one log on every process with the configuration for debugging.
     logging.basicConfig(
         format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
@@ -430,7 +430,10 @@ def preprocess_function(examples):
 
     # We need to initialize the trackers we use, and also store our configuration
     if args.with_tracking:
-        accelerator.init_trackers("glue_no_trainer", args)
+        experiment_config = vars(args)
+        # TensorBoard cannot log Enums, need the raw value
+        experiment_config["lr_scheduler_type"] = experiment_config["lr_scheduler_type"].value
+        accelerator.init_trackers("glue_no_trainer", experiment_config)
 
     # Get the metric function
     if args.task_name is not None:
@@ -520,8 +523,8 @@ def preprocess_function(examples):
                     "accuracy" if args.task_name is not None else "glue": eval_metric,
                     "train_loss": total_loss,
                     "epoch": epoch,
+                    "step": completed_steps,
                 },
-                step=completed_steps,
             )
 
         if args.push_to_hub and epoch < args.num_train_epochs - 1:
diff --git a/examples/pytorch/token-classification/run_ner_no_trainer.py b/examples/pytorch/token-classification/run_ner_no_trainer.py
index 26f1ff414017..6351d2625617 100755
--- a/examples/pytorch/token-classification/run_ner_no_trainer.py
+++ b/examples/pytorch/token-classification/run_ner_no_trainer.py
@@ -219,7 +219,7 @@ def parse_args():
     )
     parser.add_argument(
         "--with_tracking",
-        required=False,
+        action="store_true",
         help="Whether to load in all available experiment trackers from the environment and use them for logging.",
     )
     args = parser.parse_args()
@@ -246,7 +246,7 @@ def main():
 
     # Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
     # If we're using tracking, we also need to initialize it here and it will pick up all supported trackers in the environment
-    accelerator = Accelerator(log_with="all") if args.with_tracking else Accelerator()
+    accelerator = Accelerator(log_with="all", logging_dir=args.output_dir) if args.with_tracking else Accelerator()
     # Make one log on every process with the configuration for debugging.
     logging.basicConfig(
         format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
@@ -545,7 +545,10 @@ def tokenize_and_align_labels(examples):
 
     # We need to initialize the trackers we use, and also store our configuration
     if args.with_tracking:
-        accelerator.init_trackers("clm_no_trainer", args)
+        experiment_config = vars(args)
+        # TensorBoard cannot log Enums, need the raw value
+        experiment_config["lr_scheduler_type"] = experiment_config["lr_scheduler_type"].value
+        accelerator.init_trackers("ner_no_trainer", experiment_config)
 
     # Metrics
     metric = load_metric("seqeval")
@@ -676,12 +679,7 @@ def compute_metrics():
         accelerator.print(f"epoch {epoch}:", eval_metric)
         if args.with_tracking:
             accelerator.log(
-                {
-                    "seqeval": eval_metric,
-                    "train_loss": total_loss,
-                    "epoch": epoch,
-                },
-                step=completed_steps,
+                {"seqeval": eval_metric, "train_loss": total_loss, "epoch": epoch, "step": completed_steps},
             )
 
         if args.push_to_hub and epoch < args.num_train_epochs - 1:
diff --git a/examples/pytorch/translation/run_translation_no_trainer.py b/examples/pytorch/translation/run_translation_no_trainer.py
index 024d70519466..706d7637fd97 100644
--- a/examples/pytorch/translation/run_translation_no_trainer.py
+++ b/examples/pytorch/translation/run_translation_no_trainer.py
@@ -258,7 +258,7 @@ def parse_args():
     )
     parser.add_argument(
         "--with_tracking",
-        required=False,
+        action="store_true",
         help="Whether to load in all available experiment trackers from the environment and use them for logging.",
     )
     args = parser.parse_args()
@@ -287,7 +287,7 @@ def main():
 
     # Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
     # If we're using tracking, we also need to initialize it here and it will pick up all supported trackers in the environment
-    accelerator = Accelerator(log_with="all") if args.with_tracking else Accelerator()
+    accelerator = Accelerator(log_with="all", logging_dir=args.output_dir) if args.with_tracking else Accelerator()
 
     # Make one log on every process with the configuration for debugging.
     logging.basicConfig(
@@ -526,7 +526,10 @@ def preprocess_function(examples):
 
     # We need to initialize the trackers we use, and also store our configuration
     if args.with_tracking:
-        accelerator.init_trackers("translation_no_trainer", args)
+        experiment_config = vars(args)
+        # TensorBoard cannot log Enums, need the raw value
+        experiment_config["lr_scheduler_type"] = experiment_config["lr_scheduler_type"].value
+        accelerator.init_trackers("translation_no_trainer", experiment_config)
 
     metric = load_metric("sacrebleu")
 
@@ -644,12 +647,7 @@ def postprocess_text(preds, labels):
 
         if args.with_tracking:
             accelerator.log(
-                {
-                    "blue": eval_metric["score"],
-                    "train_loss": total_loss,
-                    "epoch": epoch,
-                },
-                step=completed_steps,
+                {"blue": eval_metric["score"], "train_loss": total_loss, "epoch": epoch, "step": completed_steps},
             )
 
         if args.push_to_hub and epoch < args.num_train_epochs - 1: