pytorch · rpsilva-aws · Jul 8, 2025
diff --git a/test/test_mp_collective_permute.py b/test/test_mp_collective_permute.py
@@ -5,26 +5,68 @@
 import torch_xla.core.xla_model as xm
 
 
+def _test_single_tensor_collective_permute(device, world_size, ordinal, pairs):
+  value = torch.tensor([ordinal] * 100, dtype=torch.int32, device=device)
+  result_tensor = xm.collective_permute(value, pairs)
+
+  result = result_tensor.cpu().tolist()
+  expected = [ordinal - 1] * 100 if ordinal != 0 else [world_size - 1] * 100
+
+  if result != expected:
+    print(f"Wrong result from core {ordinal}: {result}", file=sys.stderr)
+    return False
+  return True
+
+
+def _test_multi_tensor_collective_permute(device, world_size, ordinal, pairs):
+  tensor1 = torch.tensor([ordinal] * 50, dtype=torch.int32, device=device)
+  tensor2 = torch.tensor([ordinal + 100] * 75, dtype=torch.int32, device=device)
+  tensor3 = torch.tensor(
+      [ordinal + 200] * 25, dtype=torch.float32, device=device)
+
+  result_list = xm.collective_permute([tensor1, tensor2, tensor3], pairs)
+  expected_ordinal = ordinal - 1 if ordinal != 0 else world_size - 1
+
+  result1 = result_list[0].cpu().tolist()
+  expected1 = [expected_ordinal] * 50
+  if result1 != expected1:
+    print(f"Wrong result from core {ordinal}: {result1}", file=sys.stderr)
+    return False
+
+  result2 = result_list[1].cpu().tolist()
+  expected2 = [expected_ordinal + 100] * 75
+  if result2 != expected2:
+    print(f"Wrong result from core {ordinal}: {result2}", file=sys.stderr)
+    return False
+
+  result3 = result_list[2].cpu().tolist()
+  expected3 = [expected_ordinal + 200.0] * 25
+  if result3 != expected3:
+    print(f"Wrong result from core {ordinal}: {result3}", file=sys.stderr)
+    return False
+
+  return True
+
+
 def _mp_fn(index):
   device = torch_xla.device()
   if xm.xla_device_hw(device) in ['TPU', 'NEURON']:
     world_size = xr.world_size()
     ordinal = xr.global_ordinal()
-    value = torch.tensor([ordinal] * 100, dtype=torch.int32, device=device)
     pairs = []
     for i in range(1, world_size):
       pairs.append([i - 1, i])
     pairs.append([world_size - 1, 0])
-    result_tensor = xm.collective_permute(value, pairs)
-
-    result = result_tensor.cpu().tolist()
-    expected = [ordinal - 1] * 100 if ordinal != 0 else [world_size - 1] * 100
-
-    if result != expected:
-      print(f"Wrong result from core {ordinal}: {result}", file=sys.stderr)
+    if not _test_single_tensor_collective_permute(device, world_size, ordinal,
+                                                  pairs):
+      sys.exit(1)
+    if not _test_multi_tensor_collective_permute(device, world_size, ordinal,
+                                                 pairs):
       sys.exit(1)
   else:
-    print(f"Default device {device} is not a supported device", file=sys.stderr)
+    print(
+        f"Device {device} is not a supported device for this test",
+        file=sys.stderr)
 
 
 if __name__ == '__main__':

diff --git a/torch_xla/core/xla_model.py b/torch_xla/core/xla_model.py
@@ -744,8 +744,9 @@ def all_to_all(value: torch.Tensor,
   return result[0]
 
 
-def collective_permute(value: torch.Tensor,
-                       pairs: List[List[int]]) -> torch.Tensor:
+def collective_permute(
+    tensors: Union[torch.Tensor, List[torch.Tensor]],
+    pairs: List[List[int]]) -> Union[torch.Tensor, List[torch.Tensor]]:
   """Performs a XLA `CollectivePermute()` operation on the input tensor.
 
   WARNING: This function is not very reliable, may produce wrong results under
@@ -754,20 +755,25 @@ def collective_permute(value: torch.Tensor,
   See: https://www.tensorflow.org/xla/operation_semantics#collectivepermute
 
   Args:
-    value (torch.Tensor): The input tensor.
+    tensors: Either a single `torch.Tensor` or a list of `torch.Tensor` to
+      perform the collective permute over.
     pairs (list): A list of (source_replica_id, target_replica_id) pairs,
       representing the sender and receiver for the `collective_permute()`
       operation. Example: `[[0, 1], [1, 2], [2, 0]]` defines three pairs. The
         tensor will be sent from replica 0 to replica 1, replica 1 to replica 2,
         and replica 2 to replica 0.
 
   Returns:
-    The result `torch.Tensor` of the `collective_permute()` operation.
+    A single or list of `torch.Tensor` results of the `collective_permute()` operation.
   """
+  is_single_operand = isinstance(tensors, torch.Tensor)
+  assert is_single_operand or (isinstance(tensors, list) and all(
+      isinstance(v, torch.Tensor) for v in tensors))
+
   token, devctx = _get_all_reduce_token()
-  result = torch_xla._XLAC._xla_collective_permute(value, token, pairs)
-  torch_xla._XLAC._set_all_reduce_token(devctx.device, result[1])
-  return result[0]
+  result = torch_xla._XLAC._xla_collective_permute(tensors, token, pairs)
+  torch_xla._XLAC._set_all_reduce_token(devctx.device, result[-1])
+  return result[0] if is_single_operand else result[:-1]
 
 
 def collective_broadcast(tensors: List[torch.Tensor],

diff --git a/torch_xla/csrc/cross_replica_reduces.cpp b/torch_xla/csrc/cross_replica_reduces.cpp
@@ -378,6 +378,24 @@ CollectivePermuteResult BuildCollectivePermute(
   return {result, token_handler.GetNewToken(result)};
 }
 
+MultiCollectivePermuteResult BuildCollectivePermute(
+    absl::Span<const xla::XlaOp> inputs, xla::XlaOp token,
+    const std::vector<std::pair<int64_t, int64_t>>& source_target_pairs) {
+  TokenHandler token_handler(token);
+  std::vector<xla::XlaOp> result(inputs.size());
+  std::vector<xla::XlaOp> input_ops;
+  for (const auto& input : inputs) {
+    const xla::Shape& input_shape = ShapeHelper::ShapeOfXlaOp(input);
+    input_ops.push_back(token_handler.GetInput(input, &input_shape));
+  }
+  xla::XlaOp collective_result =
+      xla::MultiCollectivePermute(input_ops, source_target_pairs);
+  for (size_t i = 0; i < inputs.size(); ++i) {
+    result[i] = xla::GetTupleElement(collective_result, i);
+  }
+  return {result, token_handler.GetNewToken(result[0])};
+}
+
 SendResult BuildSendWithToken(xla::XlaOp input, xla::XlaOp token,
                               int64_t channel_id) {
   xla::ChannelHandle channel_handle;

diff --git a/torch_xla/csrc/cross_replica_reduces.h b/torch_xla/csrc/cross_replica_reduces.h
@@ -60,6 +60,11 @@ struct ReduceScatterResultCoalesced {
   xla::XlaOp token;
 };
 
+struct MultiCollectivePermuteResult {
+  std::vector<xla::XlaOp> results;
+  xla::XlaOp token;
+};
+
 std::vector<xla::XlaOp> BuildAllReduce(
     AllReduceType reduce_type, absl::Span<const xla::XlaOp> operands,
     xla::XlaOp token, double scale,
@@ -90,6 +95,10 @@ CollectivePermuteResult BuildCollectivePermute(
     xla::XlaOp input, xla::XlaOp token,
     const std::vector<std::pair<int64_t, int64_t>>& source_target_pairs);
 
+MultiCollectivePermuteResult BuildCollectivePermute(
+    absl::Span<const xla::XlaOp> inputs, xla::XlaOp token,
+    const std::vector<std::pair<int64_t, int64_t>>& source_target_pairs);
+
 SendResult BuildSendWithToken(xla::XlaOp input, xla::XlaOp token,
                               int64_t channel_id);
 

diff --git a/torch_xla/csrc/init_python_bindings.cpp b/torch_xla/csrc/init_python_bindings.cpp
@@ -601,9 +601,8 @@ std::pair<at::Tensor, std::shared_ptr<torch::lazy::Value>> AllToAll(
   std::tie(result, new_token) = tensor_methods::all_to_all(
       bridge::GetXlaTensor(input), *token, split_dimension, concat_dimension,
       split_count, replica_groups, pin_layout);
-  return std::pair<at::Tensor, std::shared_ptr<torch::lazy::Value>>(
-      bridge::AtenFromXlaTensor(std::move(result)),
-      std::make_shared<torch::lazy::Value>(new_token));
+  return {bridge::AtenFromXlaTensor(std::move(result)),
+          std::make_shared<torch::lazy::Value>(new_token)};
 }
 
 std::pair<at::Tensor, std::shared_ptr<torch::lazy::Value>> CollectivePermute(
@@ -618,6 +617,24 @@ std::pair<at::Tensor, std::shared_ptr<torch::lazy::Value>> CollectivePermute(
       std::make_shared<torch::lazy::Value>(new_token));
 }
 
+std::pair<std::vector<at::Tensor>, std::shared_ptr<torch::lazy::Value>>
+CollectivePermute(
+    const std::vector<at::Tensor>& tensors,
+    const std::shared_ptr<torch::lazy::Value>& token,
+    const std::vector<std::pair<int64_t, int64_t>>& source_target_pairs) {
+  std::vector<XLATensorPtr> xtensors =
+      GetXlaTensors(tensors, /*want_all=*/true);
+  std::vector<XLATensorPtr> results;
+  torch::lazy::Value new_token;
+  std::tie(results, new_token) =
+      tensor_methods::collective_permute(xtensors, *token, source_target_pairs);
+  std::vector<at::Tensor> aten_results;
+  for (auto& xt : results) {
+    aten_results.emplace_back(bridge::AtenFromXlaTensor(std::move(xt)));
+  }
+  return {aten_results, std::make_shared<torch::lazy::Value>(new_token)};
+}
+
 void OptimizationBarrier_(std::vector<at::Tensor>& tensors) {
   std::vector<XLATensorPtr> xtensors =
       GetXlaTensors(tensors, /*want_all=*/false);
@@ -1990,6 +2007,27 @@ void InitXlaModuleBindings(py::module m) {
             result_tuple[1] = new_token;
             return result_tuple;
            })
+      .def("_xla_collective_permute",
+           [](const std::vector<at::Tensor>& inputs,
+              const std::shared_ptr<torch::lazy::Value>& token,
+              const py::list& pairs) {
+            std::vector<std::pair<int64_t, int64_t>> source_target_pairs =
+                CreateSourceTargetPairs(pairs);
+            std::vector<at::Tensor> results;
+            std::shared_ptr<torch::lazy::Value> new_token;
+            {
+              NoGilSection nogil;
+              std::tie(results, new_token) =
+                  CollectivePermute(inputs, token, source_target_pairs);
+            }
+            auto result_list = py::list(results.size() + 1);
+            for (int i = 0; i < results.size(); ++i) {
+              result_list[i] = torch::autograd::make_variable(
+                  results[i], /*requires_grad=*/results[i].requires_grad());
+            }
+            result_list[results.size()] = new_token;
+            return result_list;
+          })
       .def("_xla_send",
            [](const at::Tensor& input,
               const std::shared_ptr<torch::lazy::Value>& token,

diff --git a/torch_xla/csrc/ops/collective_permute.cpp b/torch_xla/csrc/ops/collective_permute.cpp
@@ -31,18 +31,71 @@ CollectivePermute::CollectivePermute(
           /*num_outputs=*/2, torch::lazy::MHash(source_target_pairs)),
       source_target_pairs_(std::move(source_target_pairs)) {}
 
+CollectivePermute::CollectivePermute(
+    c10::ArrayRef<torch::lazy::Value> inputs, const torch::lazy::Value& token,
+    std::vector<std::pair<int64_t, int64_t>> source_target_pairs)
+    : XlaNode(
+          xla_collective_permute, GetOperandListWithToken(inputs, token),
+          [&]() {
+            std::vector<xla::Shape> input_shapes;
+            for (const auto& input : inputs) {
+              input_shapes.push_back(GetXlaShape(input));
+            }
+            input_shapes.push_back(GetXlaShape(token));
+            auto shape_fn =
+                [&](absl::Span<const xla::XlaOp> operands) -> xla::XlaOp {
+              std::vector<xla::XlaOp> input_ops(operands.begin(),
+                                                operands.end() - 1);
+              xla::XlaOp token_op = operands.back();
+              MultiCollectivePermuteResult result = BuildCollectivePermute(
+                  input_ops, token_op, source_target_pairs);
+              std::vector<xla::XlaOp> outputs = result.results;
+              outputs.push_back(result.token);
+              return xla::Tuple(operands[0].builder(), outputs);
+            };
+            return InferOutputShape(input_shapes, shape_fn);
+          },
+          /*num_outputs=*/inputs.size() + 1,
+          torch::lazy::MHash(source_target_pairs)),
+      source_target_pairs_(std::move(source_target_pairs)) {}
+
 torch::lazy::NodePtr CollectivePermute::Clone(
     torch::lazy::OpList operands) const {
-  return torch_xla::MakeNode<CollectivePermute>(operands.at(0), operands.at(1),
-                                                source_target_pairs_);
+  if (operands.size() > 2) {
+    std::vector<torch::lazy::Value> inputs(operands.begin(),
+                                           operands.end() - 1);
+    return torch_xla::MakeNode<CollectivePermute>(inputs, operands.back(),
+                                                  source_target_pairs_);
+  } else {
+    return torch_xla::MakeNode<CollectivePermute>(
+        operands.at(0), operands.at(1), source_target_pairs_);
+  }
 }
 
 XlaOpVector CollectivePermute::Lower(LoweringContext* loctx) const {
-  xla::XlaOp input = loctx->GetOutputOp(operand(0));
-  xla::XlaOp token = loctx->GetOutputOp(operand(1));
-  CollectivePermuteResult result =
-      BuildCollectivePermute(input, token, source_target_pairs_);
-  return ReturnOps({result.result, result.token}, loctx);
+  auto& operand_list = operands();
+  size_t operand_list_size = operand_list.size();
+  if (operand_list_size > 2) {
+    std::vector<xla::XlaOp> inputs;
+    inputs.reserve(operand_list_size);
+    for (size_t i = 0; i < operand_list_size - 1; ++i) {
+      inputs.push_back(loctx->GetOutputOp(operand(i)));
+    }
+    xla::XlaOp token = loctx->GetOutputOp(operand_list.back());
+
+    MultiCollectivePermuteResult result =
+        BuildCollectivePermute(inputs, token, source_target_pairs_);
+
+    std::vector<xla::XlaOp> outputs = result.results;
+    outputs.push_back(result.token);
+    return ReturnOps(outputs, loctx);
+  } else {
+    xla::XlaOp input = loctx->GetOutputOp(operand(0));
+    xla::XlaOp token = loctx->GetOutputOp(operand(1));
+    CollectivePermuteResult result =
+        BuildCollectivePermute(input, token, source_target_pairs_);
+    return ReturnOps({result.result, result.token}, loctx);
+  }
 }
 
 std::string CollectivePermute::ToString() const {

diff --git a/torch_xla/csrc/ops/collective_permute.h b/torch_xla/csrc/ops/collective_permute.h
@@ -12,6 +12,10 @@ class CollectivePermute : public XlaNode {
       const torch::lazy::Value& input, const torch::lazy::Value& token,
       std::vector<std::pair<int64_t, int64_t>> source_target_pairs);
 
+  CollectivePermute(
+      c10::ArrayRef<torch::lazy::Value> inputs, const torch::lazy::Value& token,
+      std::vector<std::pair<int64_t, int64_t>> source_target_pairs);
+
   std::string ToString() const override;
 
   torch::lazy::NodePtr Clone(torch::lazy::OpList operands) const override;
@@ -28,4 +32,4 @@ class CollectivePermute : public XlaNode {
 
 }  // namespace torch_xla
 
-#endif  // XLA_TORCH_XLA_CSRC_OPS_COLLECTIVE_PERMUTE_H_
+#endif  // XLA_TORCH_XLA_CSRC_OPS_COLLECTIVE_PERMUTE_H_
diff --git a/torch_xla/csrc/tensor_methods.cpp b/torch_xla/csrc/tensor_methods.cpp
@@ -564,6 +564,25 @@ std::pair<XLATensorPtr, torch::lazy::Value> collective_permute(
           torch::lazy::Value(node, 1)};
 }
 
+std::pair<std::vector<XLATensorPtr>, torch::lazy::Value> collective_permute(
+    const std::vector<XLATensorPtr>& inputs, const torch::lazy::Value& token,
+    std::vector<std::pair<int64_t, int64_t>> source_target_pairs) {
+  std::vector<torch::lazy::Value> input_values;
+  input_values.reserve(inputs.size());
+  for (const auto& input : inputs) {
+    input_values.push_back(input->GetIrValue());
+  }
+  torch::lazy::NodePtr node = torch_xla::MakeNode<CollectivePermute>(
+      input_values, token, std::move(source_target_pairs));
+
+  std::vector<XLATensorPtr> result;
+  result.reserve(inputs.size());
+  for (size_t i = 0; i < inputs.size(); ++i) {
+    result.emplace_back(inputs[i]->CreateFrom(torch::lazy::Value(node, i)));
+  }
+  return {result, torch::lazy::Value(node, inputs.size())};
+}
+
 std::vector<XLATensorPtr> custom_call(
     const std::vector<XLATensorPtr>& inputs, const std::string& target,
     const std::vector<std::vector<int64_t>>& output_shapes,

diff --git a/torch_xla/csrc/tensor_methods.h b/torch_xla/csrc/tensor_methods.h
@@ -90,6 +90,10 @@ std::pair<XLATensorPtr, torch::lazy::Value> collective_permute(
     const XLATensorPtr& input, const torch::lazy::Value& token,
     std::vector<std::pair<int64_t, int64_t>> source_target_pairs);
 
+std::pair<std::vector<XLATensorPtr>, torch::lazy::Value> collective_permute(
+    const std::vector<XLATensorPtr>& inputs, const torch::lazy::Value& token,
+    std::vector<std::pair<int64_t, int64_t>> source_target_pairs);
+
 std::vector<XLATensorPtr> custom_call(
     const std::vector<XLATensorPtr>& inputs, const std::string& target,
     const std::vector<std::vector<int64_t>>& output_shapes,