[FT Decoder] Add JIT and bug fix (#1173)

* fix&feat: gelu-->relu and add jit

* docs: update yaml

Author: gongenlei

paddlenlp/ops/CMakeLists.txt

@@ -55,7 +55,7 @@ if(WITH_ENCODER AND NOT ON_INFER)
 endif()
 
 if(WITH_DECODER)
-  list(APPEND decoder_op_files fusion_decoder_op.cc fusion_decoder_op.cu)
+  list(APPEND decoding_op_files fusion_decoder_op.cc fusion_decoder_op.cu)
 endif()
 
 if(WITH_BART)

paddlenlp/ops/faster_transformer/sample/config/decoder.sample.yaml

@@ -36,4 +36,4 @@ dropout: 0.1
 weight_sharing: True
 
 # Path of trained parameter, to make prediction
-init_from_params: base_trained_models/step_final
+init_from_params: base_trained_models/step_final

paddlenlp/ops/faster_transformer/sample/decoder_sample.py

@@ -33,9 +33,9 @@ def parse_args():
         help="Path of the config file. ")
     parser.add_argument(
         "--decoder_lib",
-        default="../../build/lib/libdecoder_op.so",
+        default="../../build/lib/libdecoding_op.so",
         type=str,
-        help="Path of libdecoder_op.so. ")
+        help="Path of libdecoding_op.so. ")
     parser.add_argument(
         "--use_fp16_decoder",
         action="store_true",

paddlenlp/ops/faster_transformer/src/CMakeLists.txt

@@ -145,7 +145,7 @@ if(ON_INFER)
     set(DEPS ${DEPS} shlwapi.lib)
   endif(NOT WIN32)
 
-  cuda_add_library(decoding_infer_op ${decoding_op_files} ${decoder_op_files} SHARED)
+  cuda_add_library(decoding_infer_op ${decoding_op_files} SHARED)
   add_dependencies(decoding_infer_op extern_${THIRD_PARTY_NAME} boost)
 
   string(REPLACE "/" ";" DEMO_PATH ${DEMO})
@@ -271,8 +271,4 @@ else(ON_INFER)
   add_library(decoding_op SHARED ${decoding_op_files})
   add_dependencies(decoding_op extern_${THIRD_PARTY_NAME} boost)
   target_link_libraries(decoding_op PRIVATE -lcublas -lcudart ${lib_link} ${ft_lib_link} -lencoder)
-
-  add_library(decoder_op SHARED ${decoder_op_files})
-  add_dependencies(decoder_op extern_${THIRD_PARTY_NAME} boost)
-  target_link_libraries(decoder_op PRIVATE -lcublas -lcudart -ldecoder ${lib_link})
 endif()

paddlenlp/ops/faster_transformer/src/fusion_decoder_op.cu

@@ -76,8 +76,13 @@ std::vector<paddle::Tensor> decoder_kernel(
   typedef typename traits_::data_t data_t_;
   typedef DecoderTransformerTraits<traits_::OpType> DecoderTraits_;
   OpenDecoder<DecoderTraits_::OpType>* decoder_;
-  decoder_ = new OpenDecoder<DecoderTraits_::OpType>(
-      batch_size_, max_seq_len_, n_head, size_per_head, memory_hidden_dim_);
+  decoder_ = new OpenDecoder<DecoderTraits_::OpType>(batch_size_,
+                                                     max_seq_len_,
+                                                     n_head,
+                                                     size_per_head,
+                                                     memory_hidden_dim_,
+                                                     true,
+                                                     ActivationType::RELU);
 
   DataType_* decoder_output = reinterpret_cast<DataType_*>(
       decoder_output_tensor.mutable_data<data_t_>());

paddlenlp/ops/faster_transformer/transformer/decoder.py

@@ -20,7 +20,8 @@
 
 from paddle.fluid.layer_helper import LayerHelper
 from paddlenlp.transformers import WordEmbedding, PositionalEmbedding, position_encoding_init
-
+from paddlenlp.utils.log import logger
+from paddlenlp.ops.ext_utils import load
 from paddlenlp.ops import transfer_param
 
 
@@ -115,6 +116,13 @@ def __init__(self,
             # Maybe it has been loadad by `ext_utils.load`
             paddle.utils.cpp_extension.load_op_meta_info_and_register_op(
                 decoder_lib)
+        else:
+            if decoder_lib is not None:
+                logger.warning(
+                    "The specified decoder_lib does not exist, and it will be built automatically."
+                )
+            load("FasterTransformer", verbose=True)
+
         super(InferTransformerDecoder, self).__init__()
         self.n_head = n_head
         self.size_per_head = size_per_head