bugfix for pp (#48353)

python/paddle/distributed/passes/auto_parallel_recompute.py

@@ -469,11 +469,26 @@ class RecomputePass(PassBase):
 
                     ckpt_ops_dict[fwd_op_id][0] = False
                     if rc_op:
+                        prior_op = main_block.ops[rc_op.idx - 1]
+                        posterior_op = rc_op
+                        prior_mesh = (
+                            self._dist_context.get_op_dist_attr_for_program(
+                                prior_op
+                            ).process_mesh
+                        )
+                        posterior_mesh = (
+                            self._dist_context.get_op_dist_attr_for_program(
+                                posterior_op
+                            ).process_mesh
+                        )
+                        # NOTE if two recompute segements across two pipeline stages
+                        # not need dependecies for it
+                        if prior_mesh == posterior_mesh:
                             insert_dependencies_for_two_ops(
                                 main_block,
                                 idx,
-                            main_block.ops[rc_op.idx - 1],
-                            rc_op,
+                                prior_op,
+                                posterior_op,
                                 self._dist_context,
                                 sync=False,
                             )