move to device in run_qa instead of accelerator

examples/question-answering/run_qa.py

@@ -376,6 +376,8 @@ def main():
         token=model_args.token,
         trust_remote_code=model_args.trust_remote_code,
     )
+    device = torch.device("hpu")
+    model = model.to(device)
 
     # Tokenizer check: this script requires a fast tokenizer.
     if not isinstance(tokenizer, PreTrainedTokenizerFast):

optimum/habana/accelerate/accelerator.py

@@ -476,9 +476,6 @@ def prepare_model(self, model: torch.nn.Module, device_placement: bool = None, e
                         "limit_all_gathers": fsdp_plugin.limit_all_gathers,
                         "device_id": torch.device("hpu", torch.hpu.current_device()),
                     }
-                    # There's issue with moving view tensors to device within FSDP class [See: https://github.com/pytorch/pytorch/issues/147321]
-                    # Due to above issue, view tensor's may lead to silent incorrent behavior, while pretending to be view they're really not
-                    model = model.to(kwargs["device_id"])
                     model = FSDP(model, **kwargs)
                     if fsdp_plugin.activation_checkpointing:
                         from torch.distributed.algorithms._checkpoint.checkpoint_wrapper import (