Add Acelerator API to ddp_tutorial (pytorch#3508)

## Motivation
- Refine the DDP tutorial to make it device-agnostic by using the
Accelerator API.

## Description
- Integrate the Accelerator API into the DDP tutorial to support
multiple accelerators and improve backend initialization.
- Align the implementation with [this
example](https://github.com/pytorch/examples/blob/main/distributed/ddp/example.py)
from the PyTorch examples repository.

## Checklist
<!--- Make sure to add `x` to all items in the following checklist: -->
- [ ] The issue that is being fixed is referred in the description (see
above "Fixes #ISSUE_NUMBER")
- [ ] Only one issue is addressed in this pull request
- [ ] Labels from the issue that this PR is fixing are added to this
pull request
- [ ] No unnecessary issues are included into this pull request.


CC: @gujinghui

cc @wconstab @osalpekar @H-Huang @kwen2501

Signed-off-by: jafraustro <[email protected]>
Co-authored-by: Chien-Chin Huang <[email protected]>

Author: jafraustro

intermediate_source/ddp_tutorial.rst

@@ -102,8 +102,12 @@ be found in
         os.environ['MASTER_ADDR'] = 'localhost'
         os.environ['MASTER_PORT'] = '12355'
 
+        # We want to be able to train our model on an `accelerator <https://pytorch.org/docs/stable/torch.html#accelerators>`__
+        # such as CUDA, MPS, MTIA, or XPU.
+        acc = torch.accelerator.current_accelerator()
+        backend = torch.distributed.get_default_backend_for_device(acc)
         # initialize the process group
-        dist.init_process_group("gloo", rank=rank, world_size=world_size)
+        dist.init_process_group(backend, rank=rank, world_size=world_size)
 
     def cleanup():
         dist.destroy_process_group()
@@ -216,8 +220,11 @@ and elasticity support, please refer to `TorchElastic <https://pytorch.org/elast
         # Use a barrier() to make sure that process 1 loads the model after process
         # 0 saves it.
         dist.barrier()
+        # We want to be able to train our model on an `accelerator <https://pytorch.org/docs/stable/torch.html#accelerators>`__
+        # such as CUDA, MPS, MTIA, or XPU.
+        acc = torch.accelerator.current_accelerator()
         # configure map_location properly
-        map_location = {'cuda:%d' % 0: 'cuda:%d' % rank}
+        map_location = {f'{acc}:0': f'{acc}:{rank}'}
         ddp_model.load_state_dict(
             torch.load(CHECKPOINT_PATH, map_location=map_location, weights_only=True))
 
@@ -295,7 +302,7 @@ either the application or the model ``forward()`` method.
 
 
     if __name__ == "__main__":
-        n_gpus = torch.cuda.device_count()
+        n_gpus = torch.accelerator.device_count()
         assert n_gpus >= 2, f"Requires at least 2 GPUs to run, but got {n_gpus}"
         world_size = n_gpus
         run_demo(demo_basic, world_size)
@@ -331,12 +338,14 @@ Let's still use the Toymodel example and create a file named ``elastic_ddp.py``.
 
 
     def demo_basic():
-        torch.cuda.set_device(int(os.environ["LOCAL_RANK"]))
-        dist.init_process_group("nccl")
+        torch.accelerator.set_device_index(int(os.environ["LOCAL_RANK"]))
+        acc = torch.accelerator.current_accelerator()
+        backend = torch.distributed.get_default_backend_for_device(acc)
+        dist.init_process_group(backend)
         rank = dist.get_rank()
         print(f"Start running basic DDP example on rank {rank}.")
         # create model and move it to GPU with id rank
-        device_id = rank % torch.cuda.device_count()
+        device_id = rank % torch.accelerator.device_count()
         model = ToyModel().to(device_id)
         ddp_model = DDP(model, device_ids=[device_id])
         loss_fn = nn.MSELoss()