Add API to assemble CPU shards to a sharded tensor (pytorch#5630)

* Add API to assemble CPU shards to a sharded tensor

* Handle replicated sharding

* Move validations into get_op_sharding

* Improve tests and error handling

test/spmd/test_xla_sharding.py

@@ -900,6 +900,139 @@ def test_op_sharding_cache(self):
     xs.mark_sharding(v, mesh, (0, None))
     self.assertEqual(met.counter_value("CreateOpSharding"), 2)
 
+  def test_from_cpu_shards_replicated(self):
+    from_cpu_shards = torch_xla._XLAC._global_tensor_from_cpu_shards
+
+    # Create an OpSharding with all devices on a single axis
+    mesh = self._get_mesh((self.n_devices,))
+    partition_spec = (None,)
+    op_sharding = mesh.get_op_sharding(partition_spec)
+    shards = [torch.arange(4)] * self.n_devices
+
+    # No shape should result in the shape of a single shard.
+    global_tensor = from_cpu_shards(shards, op_sharding)
+    self.assertTrue(torch.allclose(global_tensor.cpu(), shards[0]))
+
+    # Specify a valid shape for the global tensor
+    global_tensor = from_cpu_shards(shards, op_sharding, shards[0].shape)
+    self.assertTrue(torch.allclose(global_tensor.cpu(), shards[0]))
+
+    # All invalid shapes should raise
+    with self.assertRaises(RuntimeError):
+      from_cpu_shards(shards, op_sharding, torch.Size((5,)))
+    with self.assertRaises(RuntimeError):
+      from_cpu_shards(shards, op_sharding, torch.Size((3,)))
+    with self.assertRaises(RuntimeError):
+      from_cpu_shards(shards, op_sharding, torch.Size((2, 2)))
+
+  def test_from_cpu_shards_tiled(self):
+    from_cpu_shards = torch_xla._XLAC._global_tensor_from_cpu_shards
+
+    # Create an OpSharding with all devices on a single axis
+    mesh = self._get_mesh((self.n_devices,))
+    partition_spec = (0,)
+    op_sharding = mesh.get_op_sharding(partition_spec)
+    shards = [torch.LongTensor([i]) for i in range(self.n_devices)]
+
+    global_tensor = from_cpu_shards(shards, op_sharding)
+    self.assertTrue(
+        torch.allclose(global_tensor.cpu(), torch.arange(self.n_devices)))
+
+    # Test incorrect number of shards
+    with self.assertRaises(RuntimeError):
+      from_cpu_shards(shards[:-1], op_sharding)
+
+    # Test an invalid global shape - too many values.
+    with self.assertRaises(RuntimeError):
+      from_cpu_shards(shards, op_sharding, torch.Size((self.n_devices * 2,)))
+
+    # Test an invalid global shape - incorrect rank
+    with self.assertRaises(RuntimeError):
+      from_cpu_shards(shards, op_sharding, torch.Size((1, self.n_devices)))
+
+    # Test a valid global shape - restrict the number of meaningful values
+    # to 1, treating the rest as padding.
+    global_tensor = from_cpu_shards(shards, op_sharding, torch.Size((1,)))
+    self.assertTrue(torch.allclose(global_tensor.cpu(), torch.arange(1)))
+
+  def test_from_cpu_shards_2d(self):
+    from_cpu_shards = torch_xla._XLAC._global_tensor_from_cpu_shards
+
+    # Create an appropriate 2D mesh for the number of devices
+    if self.n_devices >= 4:
+      mesh_shape = (self.n_devices // 2, 2)
+    else:
+      mesh_shape = (1, self.n_devices)
+    mesh_2d = self._get_mesh(mesh_shape)
+
+    # Replicated sharding
+    shards = [torch.LongTensor([self.n_devices])] * self.n_devices
+    partition_spec = (None, None)
+    op_sharding = mesh_2d.get_op_sharding(partition_spec)
+    global_tensor = from_cpu_shards(shards, op_sharding)
+    self.assertTrue(torch.allclose(global_tensor.cpu(), shards[0]))
+
+    if self.n_devices > 1:
+      # Tiled sharding
+      shards = [torch.LongTensor([[i]]) for i in range(self.n_devices)]
+      partition_spec = (0, 1)
+      op_sharding = mesh_2d.get_op_sharding(partition_spec)
+      global_tensor = from_cpu_shards(shards, op_sharding)
+      expected = torch.arange(self.n_devices).reshape(2, self.n_devices // 2)
+      self.assertTrue(torch.allclose(global_tensor.cpu(), expected))
+
+      # Partially replicated sharding
+      shards = [torch.LongTensor([[i]]) for i in range(2)] * (
+          self.n_devices // 2)
+      partition_spec = (None, 1)
+      op_sharding = mesh_2d.get_op_sharding(partition_spec)
+      global_tensor = from_cpu_shards(shards, op_sharding)
+      # Partial replication along the 0th axis represents a global tensor
+      # of torch.Tensor([[0, 1]]).
+      expected = torch.arange(2).reshape(1, 2)
+      self.assertTrue(torch.allclose(global_tensor.cpu(), expected))
+
+  def test_from_cpu_shards_global_shape(self):
+    from_cpu_shards = torch_xla._XLAC._global_tensor_from_cpu_shards
+
+    mesh = self._get_mesh((self.n_devices,))
+    numel = self.n_devices**2
+    # The global tensor is torch.arange(numel).
+    shards = [
+        torch.arange(self.n_devices) + (i * self.n_devices)
+        for i in range(self.n_devices)
+    ]
+    partition_spec = (0,)
+    op_sharding = mesh.get_op_sharding(partition_spec)
+
+    # No global shape specified will include all data from the shards
+    global_tensor = from_cpu_shards(shards, op_sharding)
+    self.assertTrue(torch.allclose(global_tensor.cpu(), torch.arange(numel)))
+
+    # Too large of a global shape will error out
+    with self.assertRaises(RuntimeError):
+      from_cpu_shards(shards, op_sharding, torch.Size((numel + 1,)))
+
+    if self.n_devices > 1:
+      # When the global tensor has fewer elements than the sum of its shards,
+      # there are two cases:
+
+      #  Case 1: If the global shape is within n_devices of numel, the excess
+      #     data is treated as padding and ignored.
+      for delta in range(self.n_devices):
+        size = torch.Size((numel - delta,))
+        global_tensor = from_cpu_shards(shards, op_sharding, size)
+        expected = torch.arange(size[0])
+        self.assertTrue(torch.allclose(global_tensor.cpu(), expected))
+
+      #  Case 2: Otherwise, it is not possible to have that much padding in a
+      #     sharded tensor, and the shards are incompatible with the shape.
+      with self.assertRaises(RuntimeError):
+        shape = torch.Size((numel - self.n_devices,))
+        from_cpu_shards(shards, op_sharding, shape)
+      with self.assertRaises(RuntimeError):
+        from_cpu_shards(shards, op_sharding, torch.Size((1,)))
+
 
 if __name__ == '__main__':
   test = unittest.main()

torch_xla/csrc/init_python_bindings.cpp

@@ -34,6 +34,7 @@
 #include "torch_xla/csrc/helpers.h"
 #include "torch_xla/csrc/ir.h"
 #include "torch_xla/csrc/ir_dump_util.h"
+#include "torch_xla/csrc/layout_manager.h"
 #include "torch_xla/csrc/ops/device_data.h"
 #include "torch_xla/csrc/ops/xla_ops.h"
 #include "torch_xla/csrc/runtime/computation_client.h"
@@ -1663,6 +1664,72 @@ void InitXlaModuleBindings(py::module m) {
           }
           return std::nullopt;
         });
+  // Reassemble the CPU shards into a global tensor. A new sharded tensor is
+  // created from the local shards with the provided sharding annotation
+  // attached. The order of the shards should coincide with the order of
+  // devices returned by `torch_xla.runtime.local_runtime_devices()`.
+  m.def(
+      "_global_tensor_from_cpu_shards",
+      [](const std::vector<at::Tensor>& shards, const xla::OpSharding& sharding,
+         std::optional<std::vector<int64_t>>& global_shape) -> at::Tensor {
+        XLA_CHECK(UseVirtualDevice())
+            << "Please enable SPMD via `torch_xla.runtime.use_spmd()`";
+        auto local_devices = runtime::GetComputationClient()->GetLocalDevices();
+        XLA_CHECK(local_devices.size() == shards.size())
+            << "Must specify a shard for each local device";
+        XLA_CHECK(!global_shape.has_value() ||
+                  global_shape.value().size() == shards[0].sizes().size())
+            << "Global shape rank must agree with shard rank: expected rank "
+            << shards[0].sizes().size() << ", got "
+            << global_shape.value().size();
+
+        if (!global_shape.has_value()) {
+          // Set a default value for the global shape based on the sharding
+          // type.
+          if (sharding.type() == xla::OpSharding::OTHER) {
+            // Infer the global shape to be the shard shape scaled by the tiling
+            // dimensionality.
+            auto tile_shape = sharding.tile_assignment_dimensions();
+            global_shape = std::vector<int64_t>();
+            for (int dim = 0; dim < shards[0].sizes().size(); ++dim) {
+              auto global_dim = tile_shape[dim] * shards[0].sizes()[dim];
+              global_shape->push_back(global_dim);
+            }
+          } else if (sharding.type() == xla::OpSharding::REPLICATED) {
+            global_shape = shards[0].sizes().vec();
+          } else {
+            XLA_ERROR() << "Unsupported OpSharding type: " << sharding.type();
+          }
+        }
+
+        auto device = GetVirtualDevice();
+        auto primitive_type =
+            MakeXlaPrimitiveType(shards[0].type().scalarType(), &device);
+        xla::Shape tensor_shape = MakeArrayShapeFromDimensions(
+            global_shape.value(), /*dynamic_dimensions=*/{}, primitive_type,
+            static_cast<XlaDeviceType>(device.type()));
+        auto sharding_spec =
+            std::make_shared<XLATensor::ShardingSpec>(sharding, tensor_shape);
+
+        // Verify that the shard shape is correct for the global shape and
+        // sharding spec.
+        auto expected_shard_shape = ShardingUtil::GetShardShape(sharding_spec);
+        for (auto shard : shards) {
+          XLA_CHECK(shard.sizes() == expected_shard_shape)
+              << "Input shard shape must include padding: " << shard.sizes()
+              << " vs " << expected_shard_shape;
+        }
+
+        auto data_handle = WrapXlaData(ShardingUtil::CreateShardedData(
+            shards, local_devices, sharding_spec));
+        XLATensorPtr xla_tensor = XLATensor::Create(std::move(data_handle));
+        xla_tensor->SetShardingSpec(*sharding_spec);
+        auto tensor = bridge::AtenFromXlaTensor(std::move(xla_tensor));
+        return torch::autograd::make_variable(tensor,
+                                              shards[0].requires_grad());
+      },
+      py::arg("shards"), py::arg("sharding"),
+      py::arg("global_shape") = py::none());
   // Returns the local shards of the tensor, with values taken from the
   // underlying ComputationClient::GetDataShards. As such, the shards will
   // contain any padding that was applied to ensure they all have the same

torch_xla/csrc/xla_sharding_util.cpp

@@ -706,7 +706,8 @@ void ShardingUtil::PrepareOutputShardingPropagation(
 }
 
 runtime::ComputationClient::DataPtr ShardingUtil::CreateShardedData(
-    std::vector<at::Tensor>& local_shards, std::vector<std::string>& devices,
+    const std::vector<at::Tensor>& local_shards,
+    const std::vector<std::string>& devices,
     const XLATensor::ShardingSpecPtr& sharding_spec) {
   XLA_CHECK(local_shards.size() == devices.size())
       << "A device must be speficied for each shard";

torch_xla/csrc/xla_sharding_util.h

@@ -147,7 +147,8 @@ class ShardingUtil {
   // Transfers the individual shards to the devices and returns a DataPtr for
   // the PjRtShardedData wrapping the shards.
   static runtime::ComputationClient::DataPtr CreateShardedData(
-      std::vector<at::Tensor>& shards, std::vector<std::string>& devices,
+      const std::vector<at::Tensor>& shards,
+      const std::vector<std::string>& devices,
       const XLATensor::ShardingSpecPtr& sharding_spec);
 };
 

torch_xla/experimental/xla_sharding.py

@@ -87,6 +87,14 @@ def get_op_sharding(self,
     Return the OpSharding for the given partition spec. This is an expensive
     operation as the mesh grows, so the value is cached for reuse.
     """
+    partition_spec = _translate_named_partition_spec(self, partition_spec)
+    flat_specs = np.hstack([d for d in partition_spec])
+    specs = [d for d in flat_specs if d is not None]
+    assert all(d >= 0 and d < len(self.mesh_shape) for d in specs), \
+      f"partition_spec ({partition_spec}) contains out of bound index into mesh_shape."
+    assert len(specs) == len(np.unique(specs)), \
+    f"Each device mesh dimension should appear at most once in partition_spec {partition_spec}."
+
     tile_assignment = _get_tile_assignment(self, partition_spec)
     if len(tile_assignment.shape) > len(partition_spec):
       # Use partial replication for sharding a tensor over a higher-rank mesh
@@ -482,19 +490,12 @@ def mark_sharding(
   assert num_devices > 0, "This requires XLA supported device(s)."
   assert mesh.size() == num_devices, \
     f"{mesh.mesh_shape} is not mappable over {num_devices} devices."
-  partition_spec = _translate_named_partition_spec(mesh, partition_spec)
   # We only allow fully specified `partition_spec` to be applicable, as opposed
   # to filling in the unspecified replicated dims. Fully specified `partiion_spec`
   # should be of the same rank as `t`. This is to support partial replication
   # where the group assignment may vary with different input ranks.
   assert len(t.shape) == len(partition_spec), \
     f"Partition spec length ({len(partition_spec)}) should be equal to the input rank ({len(t.shape)})."
-  flat_specs = np.hstack([d for d in partition_spec])
-  specs = [d for d in flat_specs if d is not None]
-  assert all(d >= 0 and d < len(mesh.mesh_shape) for d in specs), \
-    f"partition_spec ({partition_spec}) contains out of bound index into mesh_shape."
-  assert len(specs) == len(np.unique(specs)), \
-    f"Each device mesh dimension should appear at most once in partition_spec {partition_spec}."
 
   op_sharding = mesh.get_op_sharding(partition_spec)