fix the bug of setting grad's placements in dtensor_to_local (#71264)

pkuzyc · web-flow · commit 3467a86eb61a · 2025-02-27T10:55:52.000+08:00
diff --git a/paddle/fluid/eager/api/manual/eager_manual/forwards/dtensor_to_local_fwd_func.cc b/paddle/fluid/eager/api/manual/eager_manual/forwards/dtensor_to_local_fwd_func.cc
@@ -57,6 +57,8 @@ paddle::Tensor dtensor_to_local_ad_function(
         ToTensorDistAttr(process_mesh, placements, input.dims());
 
     grad_node->SetGradDistAttr(grad_dist_attr);
+    grad_node->SetGradProcessMesh(process_mesh);
+    grad_node->SetGradPlacements(placements);
   }
 
   // Forward API Call
diff --git a/paddle/fluid/eager/api/manual/eager_manual/nodes/dtensor_to_local_node.cc b/paddle/fluid/eager/api/manual/eager_manual/nodes/dtensor_to_local_node.cc
@@ -78,12 +78,15 @@ DtensorToLocalGradNode::operator()(
     VLOG(3) << paddle::string::Sprintf(INPUT_PRINT_TEMPLATE, input_str);
   }
 
+  std::shared_ptr<phi::DenseTensor> grad_out_ptr =
+      std::static_pointer_cast<phi::DenseTensor>(grad_out.impl());
   // Backward call dtensor_to_local_func function
   auto dist_grad_ptr = std::make_shared<phi::distributed::DistTensor>(
-      grad_out.dims(), grad_dist_attr_);
+      grad_out_ptr,
+      out_metas[0][0].DistTensorGlobalDims(),
+      grad_process_mesh_,
+      grad_placements_);
 
-  *(dist_grad_ptr->unsafe_mutable_value()) =
-      *(static_cast<phi::DenseTensor*>(grad_out.impl().get()));
   grad_input.set_impl(dist_grad_ptr);
 
   VLOG(5) << "Finish C++ API: dtensor_to_local_func";
diff --git a/paddle/fluid/eager/api/manual/eager_manual/nodes/nodes.h b/paddle/fluid/eager/api/manual/eager_manual/nodes/nodes.h
@@ -493,11 +493,21 @@ class DtensorToLocalGradNode : public egr::GradNodeBase {
     grad_dist_attr_ = dist_attr;
   }
 
+  void SetGradPlacements(const phi::distributed::Placements& placements) {
+    grad_placements_ = placements;
+  }
+
+  void SetGradProcessMesh(const phi::distributed::ProcessMesh& process_mesh) {
+    grad_process_mesh_ = process_mesh;
+  }
+
  private:
   // TensorWrappers
   egr::TensorWrapper input_;
 
   phi::distributed::TensorDistAttr grad_dist_attr_;
+  phi::distributed::Placements grad_placements_;
+  phi::distributed::ProcessMesh grad_process_mesh_;
 };
 
 class DtensorFromLocalGradNode : public egr::GradNodeBase {
diff --git a/paddle/fluid/framework/new_executor/instruction/instruction_util.cc b/paddle/fluid/framework/new_executor/instruction/instruction_util.cc
@@ -173,7 +173,12 @@ phi::DeviceContext* ParseDeviceContext(pir::Operation* op,
             op_name.compare(paddle::dialect::AllGatherOp::name()) == 0 ||
             op_name.compare(paddle::dialect::MpAllreduceSum_Op::name()) == 0 ||
             op_name.compare(paddle::dialect::CIdentity_Op::name()) == 0 ||
-            op_name.compare(paddle::dialect::CConcatOp::name()) == 0) {
+            op_name.compare(paddle::dialect::CConcatOp::name()) == 0 ||
+            op_name.compare(paddle::dialect::CConcatOp::name()) == 0 ||
+            op_name.compare(paddle::dialect::AllGatherOp::name()) == 0 ||
+            op_name.compare(paddle::dialect::AllToAllOp::name()) == 0 ||
+            op_name.compare(
+                paddle::dialect::CSoftmaxWithCrossEntropyOp::name()) == 0) {
           if (phi::is_gpu_place(place) && execution_stream == kDefaultStream) {
             if (origin_dev_ctx != nullptr) {
               // set stream
diff --git a/paddle/fluid/framework/new_executor/interpreter/interpreter_util.cc b/paddle/fluid/framework/new_executor/interpreter/interpreter_util.cc
@@ -1523,6 +1523,7 @@ std::unordered_map<std::string, std::set<std::string>> GetNoNeedBufferValues(
           no_need_buffer_vars.insert(name);
         } else {
           no_need_buffer_vars.erase(name);
+          break;
         }
       }
     }
@@ -1535,6 +1536,7 @@ std::unordered_map<std::string, std::set<std::string>> GetNoNeedBufferValues(
             no_need_buffer_vars.insert(name);
           } else {
             no_need_buffer_vars.erase(name);
+            break;
           }
         }
       }
diff --git a/paddle/fluid/pir/dialect/distributed/ir/dist_op.h b/paddle/fluid/pir/dialect/distributed/ir/dist_op.h
@@ -108,7 +108,7 @@ class DtensorToLocalOp
   //   void VerifySig();
 };
 
-class MoESubMeshTensorsOp : public pir::Op<MoESubMeshTensorsOp> {
+class MoESubMeshTensorsOp : public pir::Op<MoESubMeshTensorsOp, VjpInterface> {
  public:
   using Op::Op;
   static const char* name() { return "dist_op.moe_sub_mesh_tensors"; }
diff --git a/python/paddle/distributed/auto_parallel/api.py b/python/paddle/distributed/auto_parallel/api.py
@@ -495,6 +495,7 @@ def _cal_global_shape(local_shape, mesh, placements):
 def moe_global_mesh_tensor(
     local_tensor_list, mesh, placements, local_mesh_dim=-1
 ):
+    placements = copy.deepcopy(placements)
     local_mesh_list, local_placements = _get_sub_meshes_and_local_placements(
         mesh, placements, local_mesh_dim
     )
@@ -548,16 +549,17 @@ def moe_global_mesh_tensor(
         global_dims = _cal_global_shape(
             local_tensor._local_shape, mesh, placements
         )
-        return paddle.jit.dy2static.py_layer.StaticPyLayer(
-            _moe_global_mesh_tensor
-        ).apply(
+        dist_tensor = paddle._C_ops.moe_global_mesh_tensor(
             local_tensor_list,
             local_mesh_list,
             local_placements,
             mesh,
             placements,
             global_dims,
         )
+        dist_tensor.stop_gradient = local_tensor_list[0].stop_gradient
+        dist_tensor.persistable = local_tensor_list[0].persistable
+        return dist_tensor
     else:
         raise NotImplementedError(
             "dtensor_from_local_list() are only supported in dynamic and pir mode."
@@ -691,6 +693,7 @@ def moe_sub_mesh_tensors(
     """
     Get the local part of the ``dist_tensor`` on the specific ``local_mesh_dim``.
     """
+    global_placements = copy.deepcopy(global_placements)
     local_mesh_list, local_placements = _get_sub_meshes_and_local_placements(
         global_mesh, global_placements, local_mesh_dim
     )
@@ -705,17 +708,17 @@ def moe_sub_mesh_tensors(
             global_placements,
         )
     elif paddle.framework.in_pir_mode():
-
-        return paddle.jit.dy2static.py_layer.StaticPyLayer(
-            _moe_sub_mesh_tensors
-        ).apply(
+        local_tensors = paddle._C_ops.moe_sub_mesh_tensors(
             dist_tensor,
             local_mesh_list,
             local_placements,
-            local_mesh_dim,
             global_mesh,
             global_placements,
         )
+        for local_tensor in local_tensors:
+            local_tensor.stop_gradient = dist_tensor.stop_gradient
+            local_tensor.persistable = dist_tensor.persistable
+        return local_tensors
     else:
         raise NotImplementedError(
             "moe_sub_mesh_tensors is only supported in dynamic mode."
diff --git a/python/paddle/distributed/auto_parallel/placement_type.py b/python/paddle/distributed/auto_parallel/placement_type.py
@@ -86,8 +86,10 @@ def to_dim_map(placements, tensor_dims):
         if placement.is_shard():
             shard_dim = cast(Shard, placement).get_dim()
             if dim_map[shard_dim] > -1:
-                raise Exception(
-                    "Tensor dim {shard_dim} is already sharded on mesh dim {dim_map[shard_dim]}"
+                import logging
+
+                logging.warning(
+                    f"Tensor dim {shard_dim} is already sharded on mesh dim {dim_map[shard_dim]}."
                 )
 
             dim_map[shard_dim] = i
diff --git a/test/auto_parallel/pir/test_moe_api.py b/test/auto_parallel/pir/test_moe_api.py
@@ -127,14 +127,23 @@ def check_results(
         local_meshes,
         local_dims_mapping,
     ):
-        # local_tensors_from_dtensor op
-        self.check_dist_attr(ops[4], local_meshes, local_dims_mapping)
-        # dtensor_from_local_list op
-        self.check_dist_attr(ops[5], [global_mesh], global_dims_mapping)
-        # grad op for dtensor_from_local_list
-        self.check_dist_attr(ops[10], local_meshes, local_dims_mapping)
-        # grad op for local_tensors_from_dtensor op
-        self.check_dist_attr(ops[11], [global_mesh], global_dims_mapping)
+        op_names = [
+            "dist_op.moe_sub_mesh_tensors",
+            "dist_op.moe_global_mesh_tensor",
+        ]
+        ops_to_check = [op for op in ops if op.name() in op_names]
+        # moe_sub_mesh_tensors op
+        self.check_dist_attr(ops_to_check[0], local_meshes, local_dims_mapping)
+        # moe_global_mesh_tensor op
+        self.check_dist_attr(
+            ops_to_check[1], [global_mesh], global_dims_mapping
+        )
+        # grad op for moe_global_mesh_tensor
+        self.check_dist_attr(ops_to_check[2], local_meshes, local_dims_mapping)
+        # grad op for moe_sub_mesh_tensors op
+        self.check_dist_attr(
+            ops_to_check[3], [global_mesh], global_dims_mapping
+        )
 
 
 if __name__ == "__main__":

Original file line number	Diff line number	Diff line change
`@@ -57,6 +57,8 @@ paddle::Tensor dtensor_to_local_ad_function(`
`57`	`57`	`ToTensorDistAttr(process_mesh, placements, input.dims());`
`58`	`58`
`59`	`59`	`grad_node->SetGradDistAttr(grad_dist_attr);`
	`60`	`+ grad_node->SetGradProcessMesh(process_mesh);`
	`61`	`+ grad_node->SetGradPlacements(placements);`
`60`	`62`	`}`
`61`	`63`
`62`	`64`	`// Forward API Call`
Original file line number	Diff line number	Diff line change
`@@ -1523,6 +1523,7 @@ std::unordered_map<std::string, std::set<std::string>> GetNoNeedBufferValues(`
`1523`	`1523`	`no_need_buffer_vars.insert(name);`
`1524`	`1524`	`} else {`
`1525`	`1525`	`no_need_buffer_vars.erase(name);`
	`1526`	`+ break;`
`1526`	`1527`	`}`
`1527`	`1528`	`}`
`1528`	`1529`	`}`
`@@ -1535,6 +1536,7 @@ std::unordered_map<std::string, std::set<std::string>> GetNoNeedBufferValues(`
`1535`	`1536`	`no_need_buffer_vars.insert(name);`
`1536`	`1537`	`} else {`
`1537`	`1538`	`no_need_buffer_vars.erase(name);`
	`1539`	`+ break;`
`1538`	`1540`	`}`
`1539`	`1541`	`}`
`1540`	`1542`	`}`