Gaudi3 CI (#38790)

tests/trainer/test_trainer.py

@@ -3062,6 +3062,7 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
     # the test slower.
     @require_torch_non_multi_accelerator
     @run_test_using_subprocess
+    @run_first
     @slow
     def test_can_resume_training_lm(self):
         # Check if it works for a simple language modeling example
@@ -3517,7 +3518,6 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
                 )
 
     @slow
-    @run_first
     def test_trainer_eval_mrpc(self):
         MODEL_ID = "google-bert/bert-base-cased-finetuned-mrpc"
         tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
@@ -3534,7 +3534,6 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
             self.assertLess(result["eval_loss"], 0.2)
 
     @slow
-    @run_first
     def test_trainer_eval_multiple(self):
         MODEL_ID = "openai-community/gpt2"
         tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
@@ -4125,6 +4124,7 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
             self.assertListEqual(trainer.optimizer.param_groups[1]["params"], no_wd_params)
 
     @slow
+    @run_first
     @require_non_hpu
     @require_torch_multi_accelerator
     def test_end_to_end_example(self):

tests/trainer/test_trainer_distributed.py

@@ -22,6 +22,7 @@ from transformers.testing_utils import (
     execute_subprocess_async,
     get_torch_dist_unique_port,
     require_torch_multi_accelerator,
+    run_first,
     torch_device,
 )
 from transformers.training_args import ParallelMode
@@ -116,6 +117,7 @@ if is_torch_available():
 
 
 class TestTrainerDistributed(TestCasePlus):
+    @run_first
     @require_torch_multi_accelerator
     def test_trainer(self):
         distributed_args = f"""--nproc_per_node={backend_device_count(torch_device)}
@@ -199,8 +201,7 @@ if __name__ == "__main__":
     model = RegressionModel()
     training_args.per_device_train_batch_size = 1
     training_args.max_steps = 1
-    training_args.accelerator_config = {
-        "dispatch_batches": False,
-    }
+    training_args.accelerator_config.dispatch_batches = False
+
     trainer = Trainer(model, training_args, train_dataset=train_dataset)
     trainer.train()

tests/trainer/test_trainer_distributed_loss.py

@@ -18,11 +18,13 @@ from transformers.testing_utils import (
     execute_subprocess_async,
     get_torch_dist_unique_port,
     require_torch_multi_accelerator,
+    run_first,
     torch_device,
 )
 
 
 class TestTrainerDistributedLoss(TestCasePlus):
+    @run_first
     @require_torch_multi_accelerator
     def test_trainer(self):
         device_count = backend_device_count(torch_device)

tests/trainer/test_trainer_distributed_worker_seed.py

@@ -18,6 +18,7 @@ from transformers.testing_utils import (
     execute_subprocess_async,
     get_torch_dist_unique_port,
     require_torch_multi_accelerator,
+    run_first,
     torch_device,
 )
 
@@ -57,6 +58,7 @@ class DummyModel(nn.Module):
 
 
 class TestTrainerDistributedWorkerSeed(TestCasePlus):
+    @run_first
     @require_torch_multi_accelerator
     def test_trainer(self):
         device_count = backend_device_count(torch_device)