Add V2 sharding support and improve partition spec handling for multichip training (#2)

* Add V2 sharding support and improve partition spec handling for multi-chip training

These changes are required to support multi-chip training for real models on the torch-xla side.

- Added V2 OpSharding support in XlaShardingSpec, which is used internally by MpLoader for parallel input loading. The original implementation only supported V1 shardings.
- Fixed environment variable parsing for CONVERT_SHLO_TO_SHARDY - previous logic treated values like "0" or "false" as truthy.
- Added logic to compute dims, reshape_dims, and transpose_perm for V2 sharding based on mesh_shape and partition_spec.

The new logic now correctly handles cases that were previously unsupported:

  case 1: mesh_shape=(2,1,1,1), partition_spec=(0,None,None,None)
           -> dims=[2,1,1,1], reshape_dims=[2,1,1,1], transpose_perm=[0,1,2,3]

  case 2: mesh_shape=(2,1,1,1), partition_spec=(0,)
          Ã-> dims=[2], reshape_dims=[2,1,1,1], transpose_perm=[0,1,2,3]

  case 3: mesh_shape=(2,4), partition_spec=(0,None)
           -> dims=[2,1,4], reshape_dims=[2,4], transpose_perm=[0,1]

* Fix formatting according to Torch-XLA style guide

---------

Co-authored-by: Het Shah <[email protected]>

Author: sshonTT

torch_xla/csrc/init_python_bindings.cpp

@@ -1534,25 +1534,22 @@ void InitXlaModuleBindings(py::module m) {
                    const py::list& replication_groups, int sharding_type,
                    bool minibatch) {
         xla::Shape global_shape =
-            CreateComputationShapeFromTensor(tensor, nullptr);
-        if (minibatch) {
-          XLA_ASSIGN_OR_THROW(
-              runtime::ComputationClient * absl_nonnull const client,
-              runtime::GetComputationClient());
-          int num_local_devices = client->GetLocalDevices().size();
-          int num_global_devices = client->GetAllDevices().size();
-          XLA_CHECK(tile_assignment.size() == num_global_devices)
-              << "Minibatch sharding only supports sharding along the batch "
-                 "dimension";
-          int batch_dim_shape =
-              tensor.sizes()[0] * num_global_devices / num_local_devices;
-          global_shape.set_dimensions(0, batch_dim_shape);
-        }
+            ShardingUtil::GetAdjustedGlobalShape(tensor, minibatch);
         return std::make_shared<XLATensor::ShardingSpec>(
             ShardingUtil::CreateOpSharding(
                 tile_assignment, group_assignment, replication_groups,
                 ShardingUtil::ShardingType(sharding_type)),
             global_shape, minibatch);
+      })
+      .def_init([](at::Tensor tensor, const py::list& dims,
+                   const py::list& reshape_dims, const py::list& transpose_perm,
+                   bool minibatch) {
+        xla::Shape global_shape =
+            ShardingUtil::GetAdjustedGlobalShape(tensor, minibatch);
+        return std::make_shared<XLATensor::ShardingSpec>(
+            ShardingUtil::CreateIotaOpSharding(dims, reshape_dims,
+                                               transpose_perm),
+            global_shape, minibatch);
       });
 
   // Define the _XLAC.IrValue class.

torch_xla/csrc/xla_sharding_util.cpp

@@ -889,4 +889,23 @@ bool ShardingUtil::GetAutoSharding() {
   }
   return use_auto_sharding;
 }
+
+xla::Shape ShardingUtil::GetAdjustedGlobalShape(const at::Tensor& tensor,
+                                                bool minibatch) {
+  xla::Shape global_shape = CreateComputationShapeFromTensor(tensor, nullptr);
+  if (minibatch) {
+    int num_local_devices =
+        runtime::GetComputationClientOrDie()->GetLocalDevices().size();
+    int num_global_devices =
+        runtime::GetComputationClientOrDie()->GetAllDevices().size();
+    XLA_CHECK(tile_assignment.size() == num_global_devices)
+        << "Minibatch sharding only supports sharding along the batch "
+           "dimension";
+    int batch_dim_shape =
+        tensor.sizes()[0] * num_global_devices / num_local_devices;
+    global_shape.set_dimensions(0, batch_dim_shape);
+  }
+  return global_shape;
+}
+
 }  // namespace torch_xla

torch_xla/csrc/xla_sharding_util.h

@@ -155,6 +155,9 @@ class ShardingUtil {
 
   static void SetAutoSharding();
   static bool GetAutoSharding();
+
+  static xla::Shape GetAdjustedGlobalShape(const at::Tensor& tensor,
+                                           bool minibatch);
 };
 
 }  // namespace torch_xla

torch_xla/distributed/spmd/xla_sharding.py

@@ -154,12 +154,10 @@ def _get_op_sharding_args(self, partition_spec: PartitionSpec):
 
   @functools.lru_cache(maxsize=None)
   def _get_op_sharding_args_v2(self, partition_spec: PartitionSpec):
-    """
-    Returns the appropriate dims, reshape_dims, and transpose_perm for the given partition spec.
-    """
     partition_spec = _translate_named_partition_spec(self, partition_spec)
     self._validate_translated_partition_spec(partition_spec)
 
+    # 1. Calculate the initial part of dims based on the partition_spec.
     dims = []
     used_axes = OrderedDict()
     for axis in partition_spec:
@@ -175,14 +173,22 @@ def _get_op_sharding_args_v2(self, partition_spec: PartitionSpec):
         dims.append(self.mesh_shape[axis])
         used_axes[axis] = True
       else:
-        # Replicated mesh axis
         dims.append(1)
 
-    transpose_perm = [k for k in used_axes.keys()]
+    # 2. If the product of dims is less than the total number of devices,
+    #    append the sizes of the unused mesh axes.
+    if math.prod(dims) < math.prod(self.mesh_shape):
+      for i in range(len(self.mesh_shape)):
+        if i not in used_axes:
+          dims.append(self.mesh_shape[i])
+
+    # 3. Calculate transpose_perm (sharded axes first, then unused axes).
+    transpose_perm = list(used_axes.keys())
     for i in range(len(self.mesh_shape)):
       if i not in used_axes:
-        dims.append(self.mesh_shape[i])
         transpose_perm.append(i)
+
+    # 4. reshape_dims is always the physical mesh shape.
     reshape_dims = list(self.mesh_shape)
 
     return dims, reshape_dims, transpose_perm
@@ -592,6 +598,11 @@ def _mark_manual_sharding(
   return wrap_as_sharded_tensor(t)
 
 
+def _use_shlo_to_shardy() -> bool:
+  return os.environ.get("CONVERT_SHLO_TO_SHARDY",
+                        "").lower() in ("1", "true", "yes")
+
+
 def enable_manual_sharding(t: Union[torch.Tensor, XLAShardedTensor],
                            partition_spec: PartitionSpec,
                            *,
@@ -716,7 +727,7 @@ def mark_sharding(t: Union[torch.Tensor, XLAShardedTensor], mesh: Mesh,
     t.shard_(NamedSharding(jmesh, P(*partition_spec)))
     return t
 
-  if os.environ.get('CONVERT_SHLO_TO_SHARDY', False):
+  if _use_shlo_to_shardy():
     op_sharding = mesh.get_op_sharding_v2(partition_spec)
   else:
     op_sharding = mesh.get_op_sharding(partition_spec)
@@ -898,6 +909,9 @@ def __post_init__(self):
     self._group_assignment, self._replication_groups = _get_group_assignment(
         self._sharding_type, tile_assignment, len(partition_spec),
         replicate_dims)
+    if _use_shlo_to_shardy():
+      self.dims, self.reshape_dims, self.transpose_dims = mesh._get_op_sharding_args_v2(
+          partition_spec)
 
   def xla_spec(self, t: torch.Tensor) -> Union['XlaShardingSpec', None]:
     """
@@ -906,6 +920,13 @@ def xla_spec(self, t: torch.Tensor) -> Union['XlaShardingSpec', None]:
     """
     if not self.can_apply(t):
       return None
+
+    if _use_shlo_to_shardy():
+      # Convert to Shardy spec if the environment variable is set.
+      return torch_xla._XLAC.XlaShardingSpec(t, self.dims, self.reshape_dims,
+                                             self.transpose_dims,
+                                             self.minibatch)
+
     return torch_xla._XLAC.XlaShardingSpec(t, self._tile_assignment,
                                            self._group_assignment,
                                            self._replication_groups,