diff --git a/examples/BuddyLeNet/CMakeLists.txt b/examples/BuddyLeNet/CMakeLists.txt
index 91d9de00d1..9ea26f00e3 100644
--- a/examples/BuddyLeNet/CMakeLists.txt
+++ b/examples/BuddyLeNet/CMakeLists.txt
@@ -60,27 +60,44 @@ target_link_directories(buddy-lenet-run PRIVATE ${LLVM_LIBRARY_DIR})
 if(NOT DEFINED BUDDY_ENABLE_PNG)
   message(FATAL_ERROR "To run LeNet inference, the png library is required. Please define BUDDY_ENABLE_PNG for CMake.")
 endif()
-set(BUDDY_LENET_LIBS LENET mlir_c_runner_utils ${PNG_LIBRARIES})
+set(BUDDY_LENET_LIBS LENET mlir_c_runner_utils mlir_async_runtime mlir_runner_utils mlir_cuda_runtime ${PNG_LIBRARIES})
 
 target_link_libraries(buddy-lenet-run ${BUDDY_LENET_LIBS})
 
 set(ONE_SHOT_BUFFERIZE_OPTION "bufferize-function-boundaries=1 function-boundary-type-conversion=identity-layout-map")
 set(LOWER_TO_NVVM_OPTION "cubin-chip=sm_80 cubin-features=+ptx71 cubin-format=fatbin")
+set(CONVERT_MEMCPY_TO_GPU_OPTION "process-args=1")
+set(CONVERT_MEMCPY_TO_GPU_OPTION_DISABLE_PROCESS_ARG "process-args=0")
+
+add_custom_command(
+  OUTPUT forward_gpu.o
+  COMMAND ${BUDDY_BINARY_DIR}/buddy-opt ${BUDDY_EXAMPLES_DIR}/BuddyLeNet/forward.mlir 
+          -buffer-deallocation
+          -canonicalize -cse -expand-strided-metadata -convert-memcpy-to-gpu -gpu-async-region |
+          ${LLVM_TOOLS_BINARY_DIR}/mlir-opt -llvm-request-c-wrappers --gpu-to-llvm |
+          ${LLVM_TOOLS_BINARY_DIR}/mlir-translate -mlir-to-llvmir |
+          ${LLVM_TOOLS_BINARY_DIR}/llvm-as |
+          ${LLVM_TOOLS_BINARY_DIR}/llc -filetype=obj  -relocation-model=pic -O0 -o ${BUDDY_BINARY_DIR}/../examples/BuddyLeNet/forward_gpu.o
+  DEPENDS ${BUDDY_EXAMPLES_DIR}/BuddyLeNet/forward.mlir
+  COMMENT "Building forward_gpu.o"
+  VERBATIM)
+
 add_custom_command(
   OUTPUT subgraph0_gpu.o
   COMMAND ${LLVM_TOOLS_BINARY_DIR}/mlir-opt ${BUDDY_EXAMPLES_DIR}/BuddyLeNet/subgraph0.mlir 
             -pass-pipeline "builtin.module(func.func(tosa-to-linalg-named, tosa-to-linalg, tosa-to-tensor, tosa-to-arith))" |
-          ${LLVM_TOOLS_BINARY_DIR}/mlir-opt
-          -one-shot-bufferize=${ONE_SHOT_BUFFERIZE_OPTION}
-          -buffer-deallocation
+          ${BUDDY_BINARY_DIR}/buddy-opt
+          -one-shot-bufferize
+          -func-bufferize-dynamic-offset
           -convert-linalg-to-parallel-loops
           -canonicalize
           -gpu-map-parallel-loops
           -convert-parallel-loops-to-gpu
           -gpu-kernel-outlining
+          -buffer-deallocation
           -canonicalize
           -cse |
-          ${BUDDY_BINARY_DIR}/buddy-opt -convert-memcpy-to-gpu -gpu-async-region -canonicalize |
+          ${BUDDY_BINARY_DIR}/buddy-opt -convert-memcpy-to-gpu=${CONVERT_MEMCPY_TO_GPU_OPTION_DISABLE_PROCESS_ARG} -gpu-async-region -canonicalize |
           ${LLVM_TOOLS_BINARY_DIR}/mlir-opt -llvm-request-c-wrappers --test-lower-to-nvvm=${LOWER_TO_NVVM_OPTION} |
           ${LLVM_TOOLS_BINARY_DIR}/mlir-translate -mlir-to-llvmir |
           ${LLVM_TOOLS_BINARY_DIR}/llvm-as |
@@ -89,7 +106,7 @@ add_custom_command(
   COMMENT "Building subgraph0_gpu.o"
   VERBATIM)
 
-add_library(LENET_GPU STATIC subgraph0_gpu.o forward.o)
+add_library(LENET_GPU STATIC subgraph0_gpu.o forward_gpu.o)
 
 SET_TARGET_PROPERTIES(LENET_GPU PROPERTIES LINKER_LANGUAGE C)