From 3ec0bcbbb8736fc5b549475714fa9f8db4d7c93b Mon Sep 17 00:00:00 2001
From: Wilber <jiweibo@baidu.com>
Date: Wed, 19 Aug 2020 10:22:14 +0800
Subject: [PATCH] [Bug] Fix prune for save_inference_model about transformer
 (#25347)

---
 paddle/fluid/framework/prune.cc      | 17 +++++++++++++++
 paddle/fluid/framework/prune_test.cc | 31 ++++++++++++++++++++++++++++
 python/paddle/fluid/layers/rnn.py    |  3 ++-
 3 files changed, 50 insertions(+), 1 deletion(-)

diff --git a/paddle/fluid/framework/prune.cc b/paddle/fluid/framework/prune.cc
index 919378c929..274b0ca0d9 100644
--- a/paddle/fluid/framework/prune.cc
+++ b/paddle/fluid/framework/prune.cc
@@ -210,6 +210,23 @@ void prune_impl(const proto::ProgramDesc& input, proto::ProgramDesc* output,
       should_run.push_back(true);
     } else {
       should_run.push_back(false);
+      // If the output of an op modifies feed vars, the op should not clip.
+      // For example, in the transformer structure, the third parameter returned
+      // by beam_search op is generally assigned to a feed var. Cutting the
+      // assign op will cause an error.
+      if (parent_block_id != -1) {
+        bool flag = false;
+        for (auto& var : op_desc.outputs()) {
+          for (auto& argu : var.arguments()) {
+            if (feed_var_names.count(argu)) {
+              flag = true;
+            }
+          }
+        }
+        if (flag) {
+          should_run.back() = true;
+        }
+      }
     }
   }
 
diff --git a/paddle/fluid/framework/prune_test.cc b/paddle/fluid/framework/prune_test.cc
index eb5c241a83..12fa0c61f8 100644
--- a/paddle/fluid/framework/prune_test.cc
+++ b/paddle/fluid/framework/prune_test.cc
@@ -185,3 +185,34 @@ TEST(Prune, recurrrent_op) {
   EXPECT_EQ(pruned.blocks(0).ops_size(), 2);
   EXPECT_EQ(pruned.blocks(1).ops_size(), 1);
 }
+
+// If the output of an op modifies feed vars, the op should not clip.
+TEST(Prune, recurrrent_op_2) {
+  f::ProgramDesc program;
+  f::BlockDesc *block = program.MutableBlock(0);
+  f::BlockDesc *sub_block = program.AppendBlock(*block);
+  AddOp("one_two", {{"input", {"a"}}}, {{"output", {"b", "c"}}},
+        f::AttributeMap{}, block);
+
+  std::vector<std::string> state_var_name(1, "y");
+  AddOp("recurrent", {{"input", {"b", "c"}}}, {{"output", {"b1, c1"}}},
+        {{"ex_states", state_var_name},
+         {"states", state_var_name},
+         {"sub_block", sub_block}},
+        block);
+
+  EXPECT_TRUE(sub_block != nullptr);
+  AddOp("rnn_memory_helper", {{"input", {"x"}}}, {{"output", {"a"}}},
+        f::AttributeMap{}, sub_block);
+
+  f::proto::ProgramDesc *pdesc = program.Proto();
+  pdesc->mutable_blocks(0)->mutable_ops(1)->set_is_target(true);
+
+  f::proto::ProgramDesc pruned;
+  std::set<std::string> feed_var_names = {"x", "a"};
+
+  f::Prune(*pdesc, feed_var_names, &pruned);
+  EXPECT_EQ(pruned.blocks_size(), 2);
+  EXPECT_EQ(pruned.blocks(0).ops_size(), 2);
+  EXPECT_EQ(pruned.blocks(1).ops_size(), 1);
+}
diff --git a/python/paddle/fluid/layers/rnn.py b/python/paddle/fluid/layers/rnn.py
index 39d25a9c7f..4ec0770aaf 100644
--- a/python/paddle/fluid/layers/rnn.py
+++ b/python/paddle/fluid/layers/rnn.py
@@ -3102,7 +3102,8 @@ def beam_search_decode(ids, scores, beam_size, end_id, name=None):
                              'beam_search_encode')
     helper = LayerHelper('beam_search_decode', **locals())
     sentence_ids = helper.create_variable_for_type_inference(dtype=ids.dtype)
-    sentence_scores = helper.create_variable_for_type_inference(dtype=ids.dtype)
+    sentence_scores = helper.create_variable_for_type_inference(
+        dtype=scores.dtype)
 
     helper.append_op(
         type="beam_search_decode",
-- 
GitLab