support layer_norm fp16 in dygraph amp (#30430)

* support layer_norm fp16 in dygraph amp

* add ut

* refine code

paddle/fluid/imperative/amp_auto_cast.cc

@@ -14,6 +14,7 @@
 
 #include "paddle/fluid/imperative/amp_auto_cast.h"
 
+#include <algorithm>
 #include <memory>
 #include <string>
 #include <utility>
@@ -35,14 +36,29 @@ AmpOperators& AmpOperators::Instance() {
   return instance;
 }
 
-std::shared_ptr<std::unordered_set<std::string>> AmpOperators::GetAllowOps() {
+std::shared_ptr<std::unordered_set<std::string>>
+AmpOperators::GetMutableAllowOps() {
   return allow_ops_;
 }
 
-std::shared_ptr<std::unordered_set<std::string>> AmpOperators::GetBlockOps() {
+std::shared_ptr<std::unordered_set<std::string>>
+AmpOperators::GetMutableBlockOps() {
   return block_ops_;
 }
 
+std::ostream& operator<<(std::ostream& os, AmpOperators& ops) {
+  os << "allow ops: ";
+  auto allow_ops = ops.GetMutableAllowOps();
+  std::copy((*allow_ops).begin(), (*allow_ops).end(),
+            std::ostream_iterator<std::string>(os, " "));
+  os << "; ";
+  os << "block ops: ";
+  auto block_ops = ops.GetMutableBlockOps();
+  std::copy((*block_ops).begin(), (*block_ops).end(),
+            std::ostream_iterator<std::string>(os, " "));
+  return os;
+}
+
 inline std::string GetDtypeStr(
     const std::shared_ptr<imperative::VarBase>& var) {
   return framework::DataTypeToString(var->DataType());
@@ -115,51 +131,50 @@ static inline framework::proto::VarType::Type GetPromoteType(
 
 NameVarBaseMap AutoCastInputs(const std::string& op_type,
                               const NameVarBaseMap& ins) {
-  NameVarBaseMap new_ins = {};
-  if (AmpOperators::Instance().GetAllowOps()->count(op_type)) {
-    for (const auto& pair : ins) {
+  NameVarBaseMap new_ins(ins);
+  if (AmpOperators::Instance().GetMutableAllowOps()->count(op_type)) {
+    for (auto& pair : new_ins) {
+      // NOTE(zhiqiu): batch_norm and layer_norm support only input x is fp16.
+      if ((op_type == "batch_norm" || op_type == "layer_norm") &&
+          pair.first != "X") {
+        continue;
+      }
+
       VLOG(5) << "Op(" << op_type << "): Cast " << pair.first << " from "
               << GetDtypeStr(*pair.second.cbegin()) << " to float16";
-      for (const auto& var : pair.second) {
-        auto new_var = CastToFP16(var);
-        new_ins[pair.first].emplace_back(new_var);
+      for (auto& var : pair.second) {
+        var = CastToFP16(var);
       }
     }
     return new_ins;
-  } else if (AmpOperators::Instance().GetBlockOps()->count(op_type)) {
-    for (const auto& pair : ins) {
+  } else if (AmpOperators::Instance().GetMutableBlockOps()->count(op_type)) {
+    for (auto& pair : new_ins) {
       VLOG(5) << "Op(" << op_type << "): Cast " << pair.first << " from "
               << GetDtypeStr(*pair.second.cbegin()) << " to float";
-      for (const auto& var : pair.second) {
-        auto new_var = CastToFP32(var);
-        new_ins[pair.first].emplace_back(new_var);
+      for (auto& var : pair.second) {
+        var = CastToFP32(var);
       }
     }
     return new_ins;
   } else {
     auto dst_type = GetPromoteType(ins);
-
-    for (const auto& pair : ins) {
+    for (auto& pair : new_ins) {
+      // NOTE(zhiqiu): batch_norm and layer_norm support only input x is fp16.
+      if ((op_type == "batch_norm" || op_type == "layer_norm") &&
+          pair.first == "X" && dst_type == framework::proto::VarType::FP32) {
+        continue;
+      }
       VLOG(5) << "Op(" << op_type << "): Cast " << pair.first << " from "
               << GetDtypeStr(*pair.second.cbegin()) << " to "
               << framework::DataTypeToString(dst_type);
-      for (const auto& var : pair.second) {
-        // NOTE(zhiqiu): Conv + BN always occur together, we needn't
-        // cast X of batch_norm to FP32, which is produced by conv as FP16 type.
-        if (op_type == "batch_norm" && pair.first == "X" &&
-            dst_type == framework::proto::VarType::FP32) {
-          new_ins[pair.first].emplace_back(var);
-          continue;
-        }
-        auto new_var = dst_type == framework::proto::VarType::FP32
-                           ? CastToFP32(var)
-                           : CastToFP16(var);
-        new_ins[pair.first].emplace_back(new_var);
+      for (auto& var : pair.second) {
+        var = (dst_type == framework::proto::VarType::FP32 ? CastToFP32(var)
+                                                           : CastToFP16(var));
       }
     }
     return new_ins;
   }
-  return ins;
+  return new_ins;
 }
 
 }  // namespace imperative

paddle/fluid/imperative/amp_auto_cast.h

@@ -36,9 +36,9 @@ class AmpOperators {
 
   static AmpOperators& Instance();
 
-  std::shared_ptr<std::unordered_set<std::string>> GetAllowOps();
+  std::shared_ptr<std::unordered_set<std::string>> GetMutableAllowOps();
 
-  std::shared_ptr<std::unordered_set<std::string>> GetBlockOps();
+  std::shared_ptr<std::unordered_set<std::string>> GetMutableBlockOps();
 
  private:
   AmpOperators();  // forbid calling default constructor
@@ -52,6 +52,8 @@ class AmpOperators {
   std::shared_ptr<std::unordered_set<std::string>> block_ops_;
 };
 
+std::ostream& operator<<(std::ostream& os, AmpOperators& ops);
+
 // NOTE(zhiqiu): AutoCastGuard is used for RAII.
 class AutoCastGuard {
  public:

paddle/fluid/pybind/imperative.cc

@@ -1257,27 +1257,30 @@ void BindImperative(py::module *m_ptr) {
            py::return_value_policy::reference)
       .def("_generate_unique_name", &imperative::Tracer::GenerateUniqueName,
            py::arg("key") = "dygraph_tmp")
-      .def(
-          "_set_amp_op_list",
-          [](imperative::Tracer &self,
-             std::unordered_set<std::string> &allow_ops,
-             std::unordered_set<std::string> &block_ops) {
-            // NOTE(zhiqiu): The automatic conversion in pybind11 between
-            // c++
-            // STL and python set/list/dict involve a copy operation that
-            // prevents pass-by-reference semantics, so it is ok to swap.
-            // The reaseon why not directly pass
-            // std::shared_ptr<std::unordered_set<std::string>>
-            // is that pybind11 forbid shared_ptr<T> where T is not custom
-            // type.
-            imperative::AmpOperators::Instance().GetAllowOps()->swap(allow_ops);
-            imperative::AmpOperators::Instance().GetBlockOps()->swap(block_ops);
-          })
+      .def("_set_amp_op_list",
+           [](imperative::Tracer &self,
+              std::unordered_set<std::string> &allow_ops,
+              std::unordered_set<std::string> &block_ops) {
+             // NOTE(zhiqiu): The automatic conversion in pybind11 between
+             // c++
+             // STL and python set/list/dict involve a copy operation that
+             // prevents pass-by-reference semantics, so it is ok to swap.
+             // The reaseon why not directly pass
+             // std::shared_ptr<std::unordered_set<std::string>>
+             // is that pybind11 forbid shared_ptr<T> where T is not custom
+             // type.
+             imperative::AmpOperators::Instance().GetMutableAllowOps()->swap(
+                 allow_ops);
+             imperative::AmpOperators::Instance().GetMutableBlockOps()->swap(
+                 block_ops);
+             VLOG(4) << "AMP operators changed, "
+                     << imperative::AmpOperators::Instance();
+           })
       .def("_get_amp_op_list",
            [](imperative::Tracer &self) {
              return std::make_tuple(
-                 *(imperative::AmpOperators::Instance().GetAllowOps()),
-                 *(imperative::AmpOperators::Instance().GetBlockOps()));
+                 *(imperative::AmpOperators::Instance().GetMutableAllowOps()),
+                 *(imperative::AmpOperators::Instance().GetMutableBlockOps()));
            })
       .def("trace",
            [](imperative::Tracer &self, const std::string &type,

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

@@ -389,5 +389,21 @@ class TestResnet(unittest.TestCase):
         self.assertTrue(np.allclose(out_fp32[0], out_amp[0], atol=1.e-2))
 
 
+class TestLayerNormFp16(unittest.TestCase):
+    r''' layer_norm and batch_norm support mixed inputs, i.e., only input x is fp16
+    and other params are fp32.
+    '''
+
+    def test_layer_norm_fp16(self):
+        if fluid.is_compiled_with_cuda():
+            with fluid.dygraph.guard(fluid.CUDAPlace(0)):
+                x = paddle.rand([2, 2, 2, 3])
+                layer_norm = paddle.nn.LayerNorm(x.shape[1:])
+                with paddle.amp.auto_cast(custom_white_list=['layer_norm']):
+                    out = layer_norm(x)
+
+                self.assertTrue(out.dtype == fluid.core.VarDesc.VarType.FP16)
+
+
 if __name__ == '__main__':
     unittest.main()