Merge pull request #16162 from wzzju/fix_nan_static_quant

Fix NaN bugs for static quantization strategy (mutil-cards train).

paddle/fluid/framework/details/inplace_op_pass.cc

@@ -16,6 +16,7 @@
 #include <algorithm>
 #include <deque>
 #include <iterator>
+#include <memory>
 #include <stack>
 #include <string>
 #include <unordered_map>
@@ -263,6 +264,10 @@ void InplacePass::WithdrawModify(const NodeSwapQueue& nodes,
 void InplacePass::TryInplaceOpInputOutput(ir::Node* op,
                                           ir::Graph* graph) const {
   VLOG(4) << "Try to inplace op " << op->Name();
+  // FIXME(liuwei1031): Graph is not aware of the existence of BlockDescs and
+  // ProgramDescs.
+  // The operations related to BlockDesc or ProgramDesc should perform on Graph
+  // or Node directly!
   PADDLE_ENFORCE(op->Op() != nullptr && op->Op()->Block() != nullptr,
                  "op_desc is nullptr");
   // some pre-requirments need to meet if the op want to inplaced.

paddle/fluid/framework/details/memory_optimize_pass.cc

@@ -24,6 +24,7 @@
 #include <sstream>
 #include <string>
 #include <type_traits>
+#include <unordered_set>
 #include <vector>
 #include "gflags/gflags.h"
 #include "paddle/fluid/framework/data_type.h"
@@ -191,6 +192,10 @@ void MemoryOptimizePass::SubGraphOptimize(OpDesc* op_desc) const {
           // immediately to make the subblock variable reuse strategy take
           // effect. Because it is a single op in graph. No need to
           // update the ir nodes.
+          // FIXME(liuwei1031): Graph is not aware of the existence of
+          // BlockDescs and ProgramDescs.
+          // The operations related to BlockDesc or ProgramDesc should perform
+          // on Graph or Node directly!
           sub_op_desc->Rename(var->Name(), cache->Name());
           if (sub_op_desc->Block() != nullptr &&
               sub_op_desc->Block()->HasVar(var->Name())) {

paddle/fluid/framework/ir/graph.h

@@ -200,7 +200,12 @@ class Graph {
   // WARN: After a series of passes, the current graph can be quite
   // different from OriginProgram. Caller shouldn't assume much from
   // the returned OriginProgram.
-  const ProgramDesc &OriginProgram() const { return program_; }
+  const ProgramDesc &OriginProgram() const {
+    LOG(WARNING) << "WARN: After a series of passes, the current graph can be "
+                    "quite different from OriginProgram. So, please avoid "
+                    "using the `OriginProgram()` method!";
+    return program_;
+  }
 
   // This method takes ownership of `node`.
   ir::Node *AddNode(ir::Node *node) {

python/paddle/fluid/compiler.py

@@ -224,12 +224,10 @@ class CompiledProgram(object):
             self._build_strategy.trainers_endpoints = tps
 
         self._persistable_vars = []
-        for block_id in range(self._program_desc.num_blocks()):
-            bdesc = self._program_desc.block(block_id)
-            self._persistable_vars.extend([
-                cpt.to_text(v.name()) for v in bdesc.all_vars()
-                if v.persistable() and v.type() != core.VarDesc.VarType.RAW
-            ])
+        for node in self._graph.nodes():
+            if node.is_var() and node.var() is not None and node.var().persistable() and \
+                    node.var().type() != core.VarDesc.VarType.RAW:
+                self._persistable_vars.append(cpt.to_text(node.name()))
 
         places = list(map(_place_obj, self._places))
 

python/paddle/fluid/contrib/slim/tests/test_quantization_pass.py

@@ -12,6 +12,7 @@
 # see the license for the specific language governing permissions and
 # limitations under the license.
 
+import os
 import unittest
 import random
 import numpy as np
@@ -25,6 +26,9 @@ from paddle.fluid.contrib.slim.quantization import ConvertToInt8Pass
 from paddle.fluid.contrib.slim.quantization import TransformForMobilePass
 from paddle.fluid import core
 
+os.environ["CUDA_VISIBLE_DEVICES"] = "0"
+os.environ["CPU_NUM"] = "1"
+
 
 def linear_fc(num):
     data = fluid.layers.data(name='image', shape=[1, 32, 32], dtype='float32')
@@ -249,7 +253,11 @@ class TestQuantizationFreezePass(unittest.TestCase):
                     marked_nodes.add(op)
             test_graph.draw('.', 'test' + dev_name + quant_type, marked_nodes)
 
-        quantized_main_program = main_graph.to_program()
+        build_strategy = fluid.BuildStrategy()
+        build_strategy.memory_optimize = False
+        build_strategy.enable_inplace = False
+        binary = fluid.CompiledProgram(main_graph.graph).with_data_parallel(
+            loss_name=loss.name, build_strategy=build_strategy)
         quantized_test_program = test_graph.to_program()
         iters = 5
         batch_size = 8
@@ -264,7 +272,7 @@ class TestQuantizationFreezePass(unittest.TestCase):
         with fluid.scope_guard(scope):
             for _ in range(iters):
                 data = next(train_reader())
-                loss_v = exe.run(program=quantized_main_program,
+                loss_v = exe.run(binary,
                                  feed=feeder.feed(data),
                                  fetch_list=[loss])
                 if not for_ci:

python/paddle/fluid/framework.py

@@ -2245,10 +2245,10 @@ class IrGraph(object):
         Notes: the `graph` cannot contain a circle.
 
         Returns:
-            set(IrNode): nodes in topology order.
+            list(IrNode): nodes in topology order.
         """
         ordered_nodes = core.topology_sort(self.graph)
-        return {IrNode(n) for n in ordered_nodes}
+        return [IrNode(n) for n in ordered_nodes]
 
     def build_adjacency_list(self):
         """
@@ -2316,7 +2316,7 @@ class IrGraph(object):
         """
         Convert the graph into a Program.
 
-        Notes: When the graph includes backward operator nodes, the
+        WARN: When the graph includes backward operator nodes, the
         conversion process may be failed. Usually, this function is
         only used to convert a test graph.