Bring back `set_epoch` for Accelerate-based dataloaders (#26850)

* Working tests!

* Fix sampler

* Fix

* Update src/transformers/trainer.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Fix check

* Clean

---------
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

src/transformers/trainer.py

@@ -200,6 +200,11 @@ if is_accelerate_available():
             save_fsdp_model,
             save_fsdp_optimizer,
         )
+    DATA_SAMPLERS = [RandomSampler]
+    if version.parse(accelerate_version) > version.parse("0.23.0"):
+        from accelerate.data_loader import SeedableRandomSampler
+
+        DATA_SAMPLERS += [SeedableRandomSampler]
 
     if is_deepspeed_available():
         from accelerate.utils import DeepSpeedSchedulerWrapper
@@ -1738,7 +1743,10 @@ class Trainer:
         if not args.ignore_data_skip:
             for epoch in range(epochs_trained):
                 sampler = get_dataloader_sampler(train_dataloader)
-                is_random_sampler = isinstance(sampler, RandomSampler)
+                sampler_kinds = [RandomSampler]
+                if version.parse(accelerate_version) > version.parse("0.23.0"):
+                    sampler_kinds.append(SeedableRandomSampler)
+                is_random_sampler = isinstance(sampler, tuple(sampler_kinds))
                 if is_torch_less_than_1_11 or not is_random_sampler:
                     # We just need to begin an iteration to create the randomization of the sampler.
                     for _ in train_dataloader:
@@ -1752,6 +1760,8 @@ class Trainer:
         total_batched_samples = 0
         for epoch in range(epochs_trained, num_train_epochs):
             epoch_iterator = train_dataloader
+            if hasattr(epoch_iterator, "set_epoch"):
+                epoch_iterator.set_epoch(epoch)
 
             # Reset the past mems state at the beginning of each epoch if necessary.
             if args.past_index >= 0: