Removed auto wrapping sharding propagation, added cached spec invalidation

Author: aws-cph

test/spmd/test_xla_dtensor_spec_conversion.py

@@ -3,9 +3,12 @@
 
 import torch
 from torch.distributed.tensor import DeviceMesh, Shard, distribute_tensor
+from torch.distributed.tensor.placement_types import Replicate
 
 import torch_xla
 import torch_xla.runtime as xr
+from torch_xla.distributed.spmd import XLAShardedTensor
+from torch_xla.distributed.spmd.xla_sharding import wrap_as_sharded_tensor
 
 import unittest
 import test_xla_sharding_base
@@ -31,7 +34,6 @@ def test_xla_to_dtensor_spec_conversion(self):
     mesh = DeviceMesh("xla", list(range(device_count)))
 
     # Test different sharding patterns
-    from torch.distributed.tensor.placement_types import Replicate
     test_cases = [
         (torch.randn(100, 50), [Shard(0)]),
         (torch.randn(100, 50), [Shard(1)]),
@@ -64,30 +66,27 @@ def test_mesh_conversion(self):
     assert converted_spec.mesh.shape == original_mesh.shape
 
   def test_spec_caching(self):
-    """Test that _spec property caches results for better performance"""
-    import time
+    """Test that _spec property caches results
+    
+    Addresses PR comment: "These sorts of tests that rely on the wall clock often lead to 
+    annoying flakes in my experience. I think it's sufficient to just test that 
+    self._cached_spec has a permanent value after the first call."
+    """
     device_count = xr.global_runtime_device_count()
     mesh = DeviceMesh("xla", list(range(device_count)))
-    tensor = torch.randn(1000,
-                         1000)  # Large tensor to make spec creation noticeable
+    tensor = torch.randn(100, 100)
     xla_tensor = distribute_tensor(tensor, mesh, [Shard(0)])
 
-    # first access should create and cache the spec
-    start_time = time.time()
+    # First access should create and cache the spec
     spec1 = xla_tensor._spec
-    first_access_time = time.time() - start_time
 
-    # should be much faster due to caching
-    start_time = time.time()
-    spec2 = xla_tensor._spec
-    second_access_time = time.time() - start_time
+    # Verify the spec is cached
+    assert xla_tensor._cached_spec is not None
+    assert xla_tensor._cached_spec is spec1
 
+    # Second access should return the cached spec
+    spec2 = xla_tensor._spec
     assert spec1 is spec2
-    print(
-        f"First access: {first_access_time:.6f}s, Second access: {second_access_time:.6f}s"
-    )
-    assert second_access_time * 10 < first_access_time, \
-        f"Cached access should be much faster: {first_access_time:.6f}s vs {second_access_time:.6f}s"
 
   def _create_test_tensor_and_mesh(self, tensor_shape, mesh_shape, placements):
     """Helper to create tensor and mesh for testing"""
@@ -114,22 +113,8 @@ def test_multi_dim_sharding_spec(self):
     assert len(spec.placements) == 2
     assert spec.mesh.ndim == 2
 
-  def test_tensor_operations_preserve_spec(self):
-    """Test that tensor operations preserve sharding metadata"""
-    xla_tensor, mesh = self._create_test_tensor_and_mesh((100, 50), (-1,),
-                                                         [Shard(0)])
-
-    result_add = xla_tensor + 1
-    result_mul = xla_tensor * 2
-    result_relu = torch.relu(xla_tensor)
-
-    for result in [result_add, result_mul, result_relu]:
-      assert hasattr(result, '_spec')
-      assert result._spec.mesh.device_type == "xla"
-
   def test_mixed_placement_spec(self):
     """Test _spec for tensors with mixed shard/replicate placements"""
-    from torch.distributed.tensor.placement_types import Replicate
     device_count = xr.global_runtime_device_count()
     if device_count < 4:
       self.skipTest("Need at least 4 devices for 2D mesh")
@@ -143,6 +128,97 @@ def test_mixed_placement_spec(self):
     assert isinstance(spec.placements[0], Shard)
     assert isinstance(spec.placements[1], Replicate)
 
+  def test_sharding_info_acquisition(self):
+    """Test that non-XLAShardedTensor can acquire sharding information
+
+    Tests case of 'elem is not an XLAShardedTensor but there exists 
+    sharding information we want to acquire'
+    """
+
+    device_count = xr.global_runtime_device_count()
+    mesh_shape = (device_count,)
+    partition_spec = (0, None)
+
+    regular_tensor = torch.randn(100, 50).to('xla')
+
+    sharded_tensor = wrap_as_sharded_tensor(
+        regular_tensor, mesh_shape=mesh_shape, partition_spec=partition_spec)
+
+    # Verify the tensor acquired the sharding information
+    assert isinstance(sharded_tensor, XLAShardedTensor)
+    assert sharded_tensor.mesh_shape == mesh_shape
+    assert sharded_tensor.partition_spec == partition_spec
+
+  def test_resharding_logic(self):
+    """
+    Tests wrap_as_sharded_tensor resharding before returning XLAShardedTensor t.
+    """
+
+    device_count = xr.global_runtime_device_count()
+    if device_count < 4:
+      self.skipTest("Need at least 4 devices for resharding test")
+
+    # Initial sharding
+    initial_mesh_shape = (device_count,)
+    initial_partition_spec = (0, None)
+    new_mesh_shape = (2, device_count // 2)
+    new_partition_spec = (0, 1)
+
+    # Create tensor and verify resharding
+    tensor = torch.randn(100, 50).to('xla')
+    sharded_tensor = wrap_as_sharded_tensor(
+        tensor,
+        mesh_shape=initial_mesh_shape,
+        partition_spec=initial_partition_spec)
+    initial_spec = sharded_tensor._spec
+
+    resharded_tensor = wrap_as_sharded_tensor(
+        sharded_tensor,
+        mesh_shape=new_mesh_shape,
+        partition_spec=new_partition_spec)
+
+    # Verify resharding worked and cache was invalidated
+    assert resharded_tensor.mesh_shape == new_mesh_shape
+    assert resharded_tensor.partition_spec == new_partition_spec
+    assert resharded_tensor._spec is not initial_spec
+
+  def test_spec_invalidation_on_resharding(self):
+    """Tests cases where the cached spec may become outdated.
+    """
+
+    device_count = xr.global_runtime_device_count()
+    if device_count < 4:
+      self.skipTest("Need at least 4 devices for resharding test")
+
+    tensor = torch.randn(100, 50).to('xla')
+    initial_mesh_shape = (device_count,)
+    initial_partition_spec = (0, None)
+    new_mesh_shape = (2, device_count // 2)
+    new_partition_spec = (0, 1)
+
+    sharded_tensor = wrap_as_sharded_tensor(
+        tensor,
+        mesh_shape=initial_mesh_shape,
+        partition_spec=initial_partition_spec)
+    initial_spec = sharded_tensor._spec
+    assert sharded_tensor._cached_spec is not None
+
+    # Changing mesh_shape / partition_spec through wrap_as_sharded_tensor invalidates cache
+    resharded_tensor = wrap_as_sharded_tensor(
+        sharded_tensor,
+        mesh_shape=new_mesh_shape,
+        partition_spec=initial_partition_spec)
+    assert resharded_tensor._spec is not initial_spec
+    assert resharded_tensor._spec.mesh.shape == new_mesh_shape
+
+    initial_spec = resharded_tensor._spec
+    resharded_tensor = wrap_as_sharded_tensor(
+        resharded_tensor,
+        mesh_shape=new_mesh_shape,
+        partition_spec=new_partition_spec)
+    assert resharded_tensor._spec is not initial_spec
+    assert resharded_tensor._spec.placements[1].dim == 1
+
 
 if __name__ == '__main__':
   test = unittest.main()

torch_xla/distributed/spmd/xla_sharded_tensor.py

@@ -10,6 +10,7 @@
 from torch.distributed.tensor._dtensor_spec import DTensorSpec, TensorMeta
 from torch.distributed.device_mesh import DeviceMesh
 from torch.distributed.tensor.placement_types import Shard, Replicate
+from torch.utils._pytree import tree_map_only
 
 
 @dataclass
@@ -115,11 +116,13 @@ def __new__(cls,
         device=elem.device,
         requires_grad=kwargs.get("requires_grad", False))
     r.global_tensor = elem.detach() if r.requires_grad else elem
-    # Store mesh and partition information for DTensor compatibility
-    if mesh_shape is not None:
-      r.mesh_shape = mesh_shape
-    if partition_spec is not None:
-      r.partition_spec = partition_spec
+
+    # Initialize mesh, partition, and spec information
+    r.mesh_shape = mesh_shape or (elem.mesh_shape if isinstance(
+        elem, XLAShardedTensor) else None)
+    r.partition_spec = partition_spec or (elem.partition_spec if isinstance(
+        elem, XLAShardedTensor) else None)
+    r._cached_spec = None
     return r
 
   # Shards on the devices are materialized/available after the lazy
@@ -176,27 +179,7 @@ def unwrap(elem):
       return elem.global_tensor if isinstance(elem, XLAShardedTensor) else elem
 
     def wrap(elem):
-      if isinstance(elem,
-                    torch.Tensor) and not isinstance(elem, XLAShardedTensor):
-        # Try to get mesh/partition info from any XLAShardedTensor in args
-        mesh_shape = None
-        partition_spec = None
-
-        def find_sharded_info(x):
-          nonlocal mesh_shape, partition_spec
-          if isinstance(x, XLAShardedTensor):
-            if hasattr(x, 'mesh_shape') and x.mesh_shape:
-              mesh_shape = x.mesh_shape
-            if hasattr(x, 'partition_spec') and x.partition_spec:
-              partition_spec = x.partition_spec
-
-        tree_map(find_sharded_info, args)
-        if kwargs:
-          tree_map(find_sharded_info, kwargs)
-
-        return XLAShardedTensor(
-            elem, mesh_shape=mesh_shape, partition_spec=partition_spec)
-      return elem
+      return XLAShardedTensor(elem) if isinstance(elem, torch.Tensor) else elem
 
     # no_dispatch is only needed if you use enable_python_mode.
     # It prevents infinite recursion.
@@ -212,11 +195,11 @@ def _spec(self):
     Convert XLA sharding information to DTensorSpec for DTensor interface compatibility.
     """
     # Return cached spec if available
-    if hasattr(self, '_cached_spec'):
+    if self._cached_spec is not None:
       return self._cached_spec
 
     # use existing mesh_shape
-    if hasattr(self, 'mesh_shape') and self.mesh_shape:
+    if self.mesh_shape is not None:
       import torch_xla.runtime as xr
       device_count = xr.global_runtime_device_count()
       device_list = list(range(device_count))
@@ -226,11 +209,9 @@ def _spec(self):
       raise ValueError("mesh_shape must be specified to create DTensorSpec")
 
     # use existing partition_spec
-    if hasattr(self, 'partition_spec') and self.partition_spec:
+    if self.partition_spec is not None:
       placements = []
-      for mesh_dim in range(
-          len(self.mesh_shape
-             ) if hasattr(self, 'mesh_shape') and self.mesh_shape else 1):
+      for mesh_dim in range(len(self.mesh_shape)):
         # find tensor dimension sharded on this mesh dimension
         tensor_dim = None
         for t_dim, m_dim in enumerate(self.partition_spec):
@@ -253,6 +234,10 @@ def _spec(self):
         mesh=mesh, placements=tuple(placements), tensor_meta=tensor_meta)
     return self._cached_spec
 
+  def invalidate_spec_cache(self):
+    """Invalidate the cached DTensorSpec."""
+    self._cached_spec = None
+
   @classmethod
   def __torch_function__(cls, func, types, args=(), kwargs=None):
     return super().__torch_function__(func, types, args, kwargs)

torch_xla/distributed/spmd/xla_sharding.py

@@ -765,14 +765,27 @@ def wrap_as_sharded_tensor(t: Union[torch.Tensor, XLAShardedTensor],
                            partition_spec=None) -> XLAShardedTensor:
   # pass along mesh and partition spec information
   if not isinstance(t, XLAShardedTensor):
+    # Create a new XLAShardedTensor
     return XLAShardedTensor(
         t, mesh_shape=mesh_shape, partition_spec=partition_spec)
-  else:
-    if mesh_shape is not None:
-      t.mesh_shape = mesh_shape
-    if partition_spec is not None:
-      t.partition_spec = partition_spec
-    return t
+
+  # Update existing XLAShardedTensor if needed
+  needs_invalidate = False
+
+  # Always set mesh_shape and partition_spec if provided
+  if mesh_shape is not None:
+    t.mesh_shape = mesh_shape
+    needs_invalidate = True
+
+  if partition_spec is not None:
+    t.partition_spec = partition_spec
+    needs_invalidate = True
+
+  # Invalidate cached spec if resharding occurred
+  if needs_invalidate:
+    t.invalidate_spec_cache()
+
+  return t
 
 
 def unwrap_sharded_tensor(