Reduce memory by using all_gather_into_tensor (#1968)

* all_gather_into_tensor

* Cleanup

* Reduce memory on non-gloo

* Fin

* Check for backend too on cpu

* CPU comment

* Change scope for performance

* Bring back zeros after remembering why

* Add comment

* Add comment

* Use empty

* Comment

src/accelerate/utils/operations.py

@@ -25,7 +25,7 @@
 from ..state import PartialState
 from .constants import TORCH_DISTRIBUTED_OPERATION_TYPES
 from .dataclasses import DistributedType, TensorInformation
-from .imports import is_torch_distributed_available, is_tpu_available
+from .imports import is_torch_distributed_available, is_torch_version, is_tpu_available
 
 
 if is_tpu_available(check_device=False):
@@ -280,16 +280,39 @@ def _tpu_gather_one(tensor):
 
 
 def _gpu_gather(tensor):
+    state = PartialState()
+    if is_torch_version(">=", "1.13"):
+        gather_op = torch.distributed.all_gather_into_tensor
+    else:
+        gather_op = torch.distributed._all_gather_base
+
     def _gpu_gather_one(tensor):
         if tensor.ndim == 0:
             tensor = tensor.clone()[None]
 
         # Can only gather contiguous tensors
         if not tensor.is_contiguous():
             tensor = tensor.contiguous()
-        output_tensors = [torch.empty_like(tensor) for _ in range(torch.distributed.get_world_size())]
-        torch.distributed.all_gather(output_tensors, tensor)
-        return torch.cat(output_tensors, dim=0)
+
+        if state.backend is not None and state.backend != "gloo":
+            # We use `empty` as `all_gather_into_tensor` slightly
+            # differs from `all_gather` for better efficiency,
+            # and we rely on the number of items in the tensor
+            # rather than its direct shape
+            output_tensors = torch.empty(
+                state.num_processes * tensor.numel(),
+                dtype=tensor.dtype,
+                device=state.device,
+            )
+            gather_op(output_tensors, tensor)
+            return output_tensors.view(-1, *tensor.size()[1:])
+        else:
+            # a backend of `None` is always CPU
+            # also gloo does not support `all_gather_into_tensor`,
+            # which will result in a larger memory overhead for the op
+            output_tensors = [torch.empty_like(tensor) for _ in range(state.num_processes)]
+            torch.distributed.all_gather(output_tensors, tensor)
+            return torch.cat(output_tensors, dim=0)
 
     return recursively_apply(_gpu_gather_one, tensor, error_on_other_type=True)