Re-organize patches (#621)

* re-organize ops patches

Author: FrostML

paddlenlp/ops/CMakeLists.txt

@@ -120,7 +120,7 @@ file(TO_NATIVE_PATH ${CMAKE_BINARY_DIR}/${THIRD_PATH}/source/${THIRD_PARTY_NAME}
 file(TO_NATIVE_PATH ${OPS_SOURCE_DIR}/patches/FasterTransformer/CMakeLists.txt cmakelists_src)
 file(TO_NATIVE_PATH ${CMAKE_BINARY_DIR}/${THIRD_PATH}/source/${THIRD_PARTY_NAME}/CMakeLists.txt cmakelists_dst)
 
-file(TO_NATIVE_PATH ${OPS_SOURCE_DIR}/patches/FasterTransformer/topk_kernels.cu topk_kernels_src)
+file(TO_NATIVE_PATH ${OPS_SOURCE_DIR}/patches/FasterTransformer/cuda/topk_kernels.cu topk_kernels_src)
 file(TO_NATIVE_PATH ${CMAKE_BINARY_DIR}/${THIRD_PATH}/source/${THIRD_PARTY_NAME}/fastertransformer/cuda/topk_kernels.cu topk_kernels_dst)
 
 # TODO(guosheng): `find` seems meeting errors missing argument to `-exec', fix it

paddlenlp/ops/faster_transformer/src/fusion_decoding_op.cc

@@ -207,7 +207,7 @@ std::vector<std::vector<int64_t>> DecodingInferShape(
   } else if (decoding_strategy == "topk_sampling" ||
              decoding_strategy == "topp_sampling") {
     output_dims = {max_len, batch_size};
-    return {output_dims, {0}, sequence_length_dims};
+    return {output_dims, {1}, sequence_length_dims};
   } else {
     PD_THROW("Not supported decoding strategy. ");
   }