Correct lstm qat test (#51499)

paddle/fluid/framework/ir/mkldnn/compute_propagate_scales_mkldnn_pass.cc

@@ -37,7 +37,7 @@ void ComputePropagateScalesMkldnnPass::GetTensorFromVector(
 void ComputePropagateScalesMkldnnPass::GetQuantInfo(
     ir::Graph* graph, StringPairMap* var_quant_scales) const {
   std::unordered_map<std::string, std::vector<float>> info_map{};
-  GetInfoFromTheFirstOp(graph, "has_quant_info", "var_quant_scales", &info_map);
+  GetInfoFromTheTmpOp(graph, "has_quant_info", "var_quant_scales", &info_map);
 
   for (auto iter = info_map.begin(); iter != info_map.end(); iter++) {
     phi::DenseTensor tensor;
@@ -510,9 +510,9 @@ void ComputePropagateScalesMkldnnPass::ApplyImpl(ir::Graph* graph) const {
   UpdateReluOutputScales(graph, &var_quant_scales);
   PropagateScales(graph, &var_quant_scales, scale_immutable_ops);
 
-  // save var_quant_scales in the first op's attr
+  // save var_quant_scales in the temporary save op's attr
   // for cpu_quantize_pass
-  SaveInfoInTheFirstOp(
+  SaveInfoInTheTmpOp(
       graph, "has_quant_info", "var_quant_scales", var_quant_scales);
 }
 

paddle/fluid/framework/ir/mkldnn/cpu_quantize_pass.cc

@@ -435,7 +435,7 @@ bool CPUQuantizePass::IsOpQuantized(const Node* node) const {
 }
 
 void CPUQuantizePass::GetQuantInfo(Graph* graph) const {
-  GetInfoFromTheFirstOp(
+  GetInfoFromTheTmpOp(
       graph, "has_quant_info", "var_quant_scales", var_quant_scales_);
 }
 
@@ -1250,6 +1250,10 @@ void CPUQuantizePass::QuantizeFusionLSTM(Graph* graph) const {
     bool is_x_unsigned{false};
     auto input_x_scale = GetScaleValueForNode(x, &is_x_unsigned);
 
+    // In the QAT process scales are prepared for only int8 data type,
+    // lstm scales should behave as input is int8 to get correct accuracy
+    is_x_unsigned = false;
+
     double input_x_shift{128.};
     if (is_x_unsigned) input_x_shift = 0.;
 

paddle/fluid/framework/ir/mkldnn/mkldnn_pass_util.h

@@ -25,7 +25,7 @@ namespace ir {
 using StringPairMap =
     std::unordered_map<std::string, std::pair<bool, phi::DenseTensor>>;
 
-static void SaveInfoInTheFirstOp(
+static void SaveInfoInTheTmpOp(
     ir::Graph* graph,
     const std::string& flag,
     const std::string& key_suffix,
@@ -33,34 +33,27 @@ static void SaveInfoInTheFirstOp(
   VLOG(3) << "save variables in the first op's attr";
 
   const std::string suffix = "_" + key_suffix + "_" + flag;
-  for (auto* op_node :
-       ir::TopologyVarientSort(*graph, static_cast<ir::SortKind>(0))) {
-    if (!op_node->IsOp() || op_node->Op()->Type() == "feed" ||
-        op_node->Op()->Type() == "fetch" ||
-        op_node->Op()->Type() == "fill_constant")
-      continue;
+  OpDesc op_desc;
+  op_desc.SetType("save");
+  auto* op_node = graph->CreateOpNode(&op_desc);
 
   op_node->Op()->SetAttr(flag, true);
   for (auto iter = info_map.begin(); iter != info_map.end(); ++iter) {
     op_node->Op()->SetAttr(iter->first + suffix, iter->second);
   }
-    break;
-  }
 }
 
-static void SaveInfoInTheFirstOp(ir::Graph* graph,
+static void SaveInfoInTheTmpOp(ir::Graph* graph,
                                const std::string& flag,
                                const std::string& key_suffix,
                                const StringPairMap& info_map) {
   VLOG(3) << "save variables in the first op's attr";
 
   const std::string suffix = "_" + key_suffix + "_" + flag;
-  for (auto* op_node :
-       ir::TopologyVarientSort(*graph, static_cast<ir::SortKind>(0))) {
-    if (!op_node->IsOp() || op_node->Op()->Type() == "feed" ||
-        op_node->Op()->Type() == "fetch" ||
-        op_node->Op()->Type() == "fill_constant")
-      continue;
+
+  OpDesc op_desc;
+  op_desc.SetType("save");
+  auto* op_node = graph->CreateOpNode(&op_desc);
 
   op_node->Op()->SetAttr(flag, true);
   for (auto iter = info_map.begin(); iter != info_map.end(); ++iter) {
@@ -70,11 +63,9 @@ static void SaveInfoInTheFirstOp(ir::Graph* graph,
                            iter->second.first);
     op_node->Op()->SetAttr(iter->first + suffix, data_v);
   }
-    break;
-  }
 }
 
-static void GetInfoFromTheFirstOp(
+static void GetInfoFromTheTmpOp(
     ir::Graph* graph,
     const std::string& flag,
     const std::string& key_suffix,
@@ -84,9 +75,7 @@ static void GetInfoFromTheFirstOp(
   const std::string suffix = "_" + key_suffix + "_" + flag;
   for (auto* op_node :
        ir::TopologyVarientSort(*graph, static_cast<ir::SortKind>(0))) {
-    if (!op_node->IsOp() || op_node->Op()->Type() == "feed" ||
-        op_node->Op()->Type() == "fetch")
-      continue;
+    if (!op_node->IsOp() || op_node->Op()->Type() != "save") continue;
 
     auto* op_desc = op_node->Op();
     if (op_desc->GetAttrIfExists<bool>(flag)) {
@@ -102,12 +91,13 @@ static void GetInfoFromTheFirstOp(
           op_desc->RemoveAttr(fake_name);
         }
       }
+      graph->RemoveNode(op_node);
       break;
     }
   }
 }
 
-static void GetInfoFromTheFirstOp(ir::Graph* graph,
+static void GetInfoFromTheTmpOp(ir::Graph* graph,
                                 const std::string& flag,
                                 const std::string& key_suffix,
                                 StringPairMap* info_map) {
@@ -117,9 +107,7 @@ static void GetInfoFromTheFirstOp(ir::Graph* graph,
   const std::string suffix_is_unsigned = suffix + unsigned_flag;
   for (auto* op_node :
        ir::TopologyVarientSort(*graph, static_cast<ir::SortKind>(0))) {
-    if (!op_node->IsOp() || op_node->Op()->Type() == "feed" ||
-        op_node->Op()->Type() == "fetch")
-      continue;
+    if (!op_node->IsOp() || op_node->Op()->Type() != "save") continue;
 
     auto* op_desc = op_node->Op();
     if (op_desc->GetAttrIfExists<bool>(flag)) {
@@ -150,6 +138,7 @@ static void GetInfoFromTheFirstOp(ir::Graph* graph,
           op_desc->RemoveAttr(vector_name);
         }
       }
+      graph->RemoveNode(op_node);
       break;
     }
   }

paddle/fluid/framework/ir/mkldnn/quant_dequant_mkldnn_pass.cc

@@ -754,9 +754,9 @@ void QuantDequantMkldnnPass::ApplyImpl(ir::Graph* graph) const {
   UpdateActivations(graph);
   RemoveCtrlVars(graph);
 
-  // save var_quant_scales in the first op's attr
+  // save var_quant_scales in the temporary save op's attr
   // for compute_propagate_scales_mkldnn_pass
-  SaveInfoInTheFirstOp(
+  SaveInfoInTheTmpOp(
       graph, "has_quant_info", "var_quant_scales", var_quant_scales);
 }
 

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -430,7 +430,6 @@ void CpuPassStrategy::EnableMkldnnInt8() {
     passes_.push_back("simplify_with_basic_ops_pass");
     passes_.push_back("quant_dequant_mkldnn_pass");
     passes_.push_back("mkldnn_placement_pass");
-    passes_.push_back("constant_folding_pass");
     passes_.push_back("squeeze2_transpose2_onednn_fuse_pass");
     passes_.push_back("layer_norm_fuse_pass");
     passes_.push_back("attention_lstm_fuse_pass");
@@ -485,6 +484,7 @@ void CpuPassStrategy::EnableMkldnnInt8() {
     passes_.push_back("quant_transpose2_dequant_onednn_fuse_pass");
     passes_.push_back("int8_scale_calculation_mkldnn_pass");
     passes_.push_back("params_quantization_mkldnn_pass");
+    passes_.push_back("constant_folding_pass");
   }
   use_mkldnn_int8_ = true;
 #else

python/paddle/static/quantization/tests/quant2_int8_lstm_model.py

@@ -116,10 +116,9 @@ class TestLstmModelPTQ(unittest.TestCase):
             config.switch_ir_optim(True)
             config.enable_mkldnn()
             config.disable_mkldnn_fc_passes()  # fc passes caused dnnl error
+            config.pass_builder().insert_pass(5, "fc_lstm_fuse_pass")
             config.set_mkldnn_cache_capacity(mkldnn_cache_capacity)
             if mode == "ptq":
-                # This pass to work properly, must be added before fc_fuse_pass
-                config.pass_builder().insert_pass(5, "fc_lstm_fuse_pass")
                 config.enable_quantizer()
                 config.quantizer_config().set_quant_data(warmup_data)
                 config.quantizer_config().set_quant_batch_size(1)
@@ -244,7 +243,7 @@ class TestLstmModelPTQ(unittest.TestCase):
         )
 
         (quant_hx_acc, quant_ctc_acc, quant_fps) = self.run_program(
-            quant_model + "_int8",
+            quant_model,
             infer_data,
             num_threads,
             mkldnn_cache_capacity,