Recompute Offload: fixed bug in memcpy (#30484) (#30517)

paddle/fluid/operators/memcpy_op.h

@@ -38,10 +38,10 @@ class MemcpyFunctor {
   void operator()(const framework::LoDTensor &lod_tensor) const {
     auto &out_tensor = *out_->GetMutable<framework::LoDTensor>();
 
-    if (dst_place_type_ == 3) {
+    if (dst_place_type_ == 2) {
       framework::TensorCopy(lod_tensor, platform::CUDAPinnedPlace(), dev_ctx_,
                             &out_tensor);
-    } else if (dst_place_type_ == 2) {
+    } else if (dst_place_type_ == 1) {
       framework::TensorCopy(lod_tensor, dev_ctx_.GetPlace(), dev_ctx_,
                             &out_tensor);
     } else {

python/paddle/fluid/optimizer.py

@@ -4773,7 +4773,7 @@ class RecomputeOptimizer(Optimizer):
         return
 
     def _insert_async_memcpy_op(self, insert_idx, src_varname, dst_varname,
-                                op_role, kind):
+                                op_role, dst_place_type):
         OP_ROLE_KEY = core.op_proto_and_checker_maker.kOpRoleAttrName()
         self.block._insert_op_without_sync(
             insert_idx,
@@ -4782,8 +4782,10 @@ class RecomputeOptimizer(Optimizer):
             outputs={
                 'Out': [self._main_program.global_block().var(dst_varname)]
             },
-            attrs={"dst_place_type": int(kind),
-                   OP_ROLE_KEY: op_role})
+            attrs={
+                "dst_place_type": int(dst_place_type),
+                OP_ROLE_KEY: op_role
+            })
 
     def _insert_fetch_op(self, idx, varname):
         assert varname in self.checkpoint_name2pinned_name, "Try to fetch {} from Pinned Memory, but it is NOT a checkpoint".format(
@@ -4791,13 +4793,13 @@ class RecomputeOptimizer(Optimizer):
 
         pinned_varname = self.checkpoint_name2pinned_name[varname]
         fetch_varname = self.checkpoint_name2fetch_name[varname]
-        self._insert_async_memcpy_op(idx, pinned_varname, fetch_varname, 1, 2)
+        self._insert_async_memcpy_op(idx, pinned_varname, fetch_varname, 1, 1)
 
     def _insert_offload_op(self, idx, varname):
         assert varname in self.checkpoint_name2pinned_name, "Try to offload {} to Pinned Memory, but it is NOT a checkpoint".format(
             varname)
         pinned_varname = self.checkpoint_name2pinned_name[varname]
-        self._insert_async_memcpy_op(idx, varname, pinned_varname, 0, 3)
+        self._insert_async_memcpy_op(idx, varname, pinned_varname, 0, 2)
 
     def _insert_sync_op(self, op_idx, checkpoint_name):
         # single stream offload no need sync 

python/paddle/fluid/tests/unittests/test_memcpy_op.py

@@ -70,7 +70,7 @@ class TestMemcpy_FillConstant(unittest.TestCase):
             type='memcpy',
             inputs={'X': gpu_var},
             outputs={'Out': pinned_var},
-            attrs={'dst_place_type': 3})
+            attrs={'dst_place_type': 2})
         place = fluid.CUDAPlace(0)
         exe = fluid.Executor(place)
         gpu_, pinned_ = exe.run(main_program,
@@ -85,7 +85,7 @@ class TestMemcpy_FillConstant(unittest.TestCase):
             type='memcpy',
             inputs={'X': pinned_var},
             outputs={'Out': gpu_var},
-            attrs={'dst_place_type': 2})
+            attrs={'dst_place_type': 1})
         place = fluid.CUDAPlace(0)
         exe = fluid.Executor(place)
         gpu_, pinned_ = exe.run(main_program,
@@ -135,7 +135,7 @@ class TestMemcpyOPError(unittest.TestCase):
             type='memcpy',
             inputs={'X': selected_row_var},
             outputs={'Out': pinned_var},
-            attrs={'dst_place_type': 3})
+            attrs={'dst_place_type': 2})
         with self.assertRaises(NotImplementedError):
             place = fluid.CUDAPlace(0)
             exe = fluid.Executor(place)