Fix

PaddlePaddle · Apr 29, 2024 · 3ac969c · 3ac969c
1 parent 8325d5d
commit 3ac969c
Show file tree

Hide file tree

Showing 8 changed files with 77 additions and 345 deletions.
diff --git a/paddle/fluid/operators/cudnn_lstm_op.cc b/paddle/fluid/operators/cudnn_lstm_op.cc
diff --git a/paddle/fluid/operators/ops_signature/cudnn_lstm_sig.cc b/paddle/fluid/operators/ops_signature/cudnn_lstm_sig.cc
diff --git a/paddle/fluid/pir/dialect/op_generator/op_infermeta_func_gen.py b/paddle/fluid/pir/dialect/op_generator/op_infermeta_func_gen.py
@@ -281,9 +281,12 @@ def GenBuildOutputsPart2(
   }}
  """
 
+    # In cudnn_lstm operator, the output weight_list_grad requires the use of optional input weight_list,
+    # so "pir::VectorType {name}" outside the "if" block.
     CREATE_OPTIONAL_INPUT_VEC_METATENSOR_TEMPLATE = """  std::vector<paddle::dialect::IrTensor> vec_ir_tensor_{name};
+  pir::VectorType {name};
   if ({name}_.impl() != nullptr) {{
-    pir::VectorType {name} = {name}_.type().dyn_cast<pir::VectorType>();
+    {name} = {name}_.type().dyn_cast<pir::VectorType>();
     for (size_t i=0; i < static_cast<size_t>({name}.size()); i++) {{
         if({name}[i].isa<paddle::dialect::DenseTensorType>()) {{
           auto {name}_type = {name}[i].dyn_cast<paddle::dialect::DenseTensorType>();

diff --git a/paddle/phi/api/yaml/backward.yaml b/paddle/phi/api/yaml/backward.yaml
@@ -519,6 +519,17 @@
     func : cross_grad
     data_type : out_grad
 
+- backward_op : cudnn_lstm_grad
+  forward: cudnn_lstm (Tensor x, Tensor init_h, Tensor init_c, Tensor w, Tensor[] weight_list, Tensor sequence_length, float dropout_prob = 0.0, bool is_bidirec = false, int hidden_size = 100, int num_layers = 1, bool is_test = false, int seed = 0) -> Tensor (out), Tensor (last_h), Tensor (last_c), Tensor (reserve), Tensor (state_out)
+  args: (Tensor x, Tensor init_h, Tensor init_c, Tensor[] weight_list, Tensor sequence_length, Tensor out, Tensor reserve, Tensor state_out, Tensor out_grad, Tensor last_h_grad, Tensor last_c_grad, float dropout_prob = 0.0, bool is_bidirec = false, int hidden_size = 100, int num_layers = 1, bool is_test = false, int seed = 0)
+  output: Tensor (x_grad), Tensor (init_h_grad), Tensor (init_c_grad), Tensor[](weight_list_grad){weight_list.size()}
+  infer_meta:
+    func: CudnnLSTMGradInferMeta
+    param : [x, init_h, init_c, weight_list]
+  kernel:
+    func: cudnn_lstm_grad
+  optional: weight_list, sequence_length, weight_list_grad
+
 - backward_op : cummax_grad
   forward : cummax(Tensor x, int axis=-1, DataType dtype = DataType::INT64) -> Tensor(out), Tensor(indices)
   args : (Tensor x, Tensor indices, Tensor out_grad, int axis, DataType dtype)