fix runtime crash when rnn model inference, test=develop (#31833)

paddle/fluid/inference/analysis/passes/memory_optimize_pass.cc

@@ -103,6 +103,7 @@ void MemoryOptimizePass::CollectVarMemorySize(
                                         "merge_lod_tensor",
                                         "equal",
                                         "sequence_pool",
+                                        "recurrent",
                                         "lod_reset"};
     for (auto* tmp : node->inputs) {
       CHECK(tmp->IsOp());

paddle/fluid/operators/recurrent_op.cc

@@ -210,9 +210,10 @@ void RecurrentOp::RunImpl(const framework::Scope &scope,
   auto *block = Attr<framework::BlockDesc *>(kStepBlock);
 
   auto *program = block->Program();
-  auto ctx = executor.Prepare(
-      *program, block->ID(), Attr<std::vector<std::string>>(
-                                 kSkipEagerDeletionVars) /*skip_ref_cnt_vars*/);
+  auto ctx = executor.Prepare(*program, block->ID(),
+                              Attr<std::vector<std::string>>(
+                                  kSkipEagerDeletionVars), /*skip_ref_cnt_vars*/
+                              true);
 
   static std::mutex mutex;
   std::lock_guard<std::mutex> lock(mutex);
@@ -255,16 +256,6 @@ void RecurrentOp::RunImpl(const framework::Scope &scope,
     // Link inside::output -> outside::output
     //   outside::output[seq_offset: seq_offset + 1] = inside::output
     executor.CreateVariables(ctx->prog_, &cur_scope, ctx->block_id_);
-    if (i > 0) {
-      LinkTensorWithCallback(scope, Outputs(kOutputs), cur_scope,
-                             Outputs(kOutputs),
-                             [&](const framework::LoDTensor &src_tensor,
-                                 framework::LoDTensor *dst_tensor) {
-                               framework::Tensor src_slice =
-                                   src_tensor.Slice(seq_offset, seq_offset + 1);
-                               dst_tensor->ShareDataWith(src_slice);
-                             });
-    }
 
     // Linked now, execute!
     executor.RunPreparedContext(ctx.get(), &cur_scope,
@@ -284,6 +275,14 @@ void RecurrentOp::RunImpl(const framework::Scope &scope,
             // early.
             framework::TensorCopy(src_tensor, place, dev_ctx, &dst_out);
           });
+    } else {
+      LinkTensorWithCallback(
+          cur_scope, Outputs(kOutputs), scope, Outputs(kOutputs),
+          [&](const framework::LoDTensor &src_tensor,
+              framework::LoDTensor *dst_tensor) {
+            auto dst_out = dst_tensor->Slice(seq_offset, seq_offset + 1);
+            framework::TensorCopy(src_tensor, place, dev_ctx, &dst_out);
+          });
     }
 
     scopes.ForwardNext();

python/paddle/nn/functional/norm.py

@@ -188,10 +188,10 @@ def batch_norm(x,
 
     if in_dygraph_mode():
         # for dygraph need tuple
-        attrs = ("momentum", momentum, "epsilon", epsilon, "data_layout",
-                 data_format, "use_mkldnn", False, "fuse_with_relu", False,
-                 "use_global_stats", use_global_stats, "trainable_statistics",
-                 trainable_statistics)
+        attrs = ("momentum", momentum, "epsilon", epsilon, "is_test",
+                 not training, "data_layout", data_format, "use_mkldnn", False,
+                 "fuse_with_relu", False, "use_global_stats", use_global_stats,
+                 "trainable_statistics", trainable_statistics)
         batch_norm_out, _, _, _, _, _ = core.ops.batch_norm(
             x, weight, bias, running_mean, running_var, mean_out, variance_out,
             *attrs)
@@ -205,6 +205,7 @@ def batch_norm(x,
     attrs = {
         "momentum": momentum,
         "epsilon": epsilon,
+        "is_test": not training,
         "data_layout": data_format,
         "use_mkldnn": False,
         "fuse_with_relu": False,