diff --git a/doc/en/build.md b/doc/en/build.md
index 1f446f1f..d4e5eb25 100644
--- a/doc/en/build.md
+++ b/doc/en/build.md
@@ -152,6 +152,7 @@ cd /Mooncake-main/build/mooncake-transfer-engine/example
 ## Advanced Compile Options
 The following options can be used during `cmake ..` to specify whether to compile certain components of Mooncake.
 - `-DUSE_CUDA=[ON|OFF]`: Enable GPU Direct RDMA and NVMe-of support
+- `-DUSE_HIP=[ON|OFF]`: Enable AMD GPU support via HIP/ROCm
 - `-DUSE_CXL=[ON|OFF]`: Enable CXL support
 - `-DWITH_STORE=[ON|OFF]`: Build Mooncake Store component
 - `-DWITH_P2P_STORE=[ON|OFF]`: Enable Golang support and build P2P Store component, require go 1.23+
diff --git a/docs/source/getting_started/build.md b/docs/source/getting_started/build.md
index d02a05bb..40535377 100644
--- a/docs/source/getting_started/build.md
+++ b/docs/source/getting_started/build.md
@@ -144,6 +144,7 @@ cd /Mooncake-main/build/mooncake-transfer-engine/example
 ## Advanced Compile Options
 The following options can be used during `cmake ..` to specify whether to compile certain components of Mooncake.
 - `-DUSE_CUDA=[ON|OFF]`: Enable GPU Direct RDMA and NVMe-of support
+- `-DUSE_HIP=[ON|OFF]`: Enable AMD GPU support via HIP/ROCm
 - `-DUSE_CXL=[ON|OFF]`: Enable CXL support
 - `-DWITH_STORE=[ON|OFF]`: Build Mooncake Store component
 - `-DWITH_P2P_STORE=[ON|OFF]`: Enable Golang support and build P2P Store component, require go 1.23+
diff --git a/mooncake-common/common.cmake b/mooncake-common/common.cmake
index a84a1078..3bf80d4d 100644
--- a/mooncake-common/common.cmake
+++ b/mooncake-common/common.cmake
@@ -99,7 +99,9 @@ if (USE_NVMEOF)
 endif()
 
 if (USE_MNNVL)
-  set(USE_CUDA ON)
+  if(NOT USE_HIP)
+    set(USE_CUDA ON)
+  endif()
   add_compile_definitions(USE_MNNVL)
   message(STATUS "Multi-Node NVLink support is enabled")
 endif()
@@ -124,24 +126,46 @@ if (USE_MUSA)
 endif()
 
 if (USE_HIP)
-  if (NOT EXISTS $ENV{ROCM_PATH})
-      if (NOT EXISTS /opt/rocm)
-          set(ROCM_PATH /usr)
-      else()
-          set(ROCM_PATH /opt/rocm)
-      endif()
-  else()
-      set(ROCM_PATH $ENV{ROCM_PATH})
-  endif()
-  add_definitions(-D__HIP_PLATFORM_AMD__)
-  add_compile_definitions(USE_HIP)
+  list(APPEND CMAKE_PREFIX_PATH "/opt/rocm/lib/cmake")
+  find_package(HIP REQUIRED)
+  include_directories(${HIP_INCLUDE_DIRS})
+  add_compile_definitions(USE_HIP __HIP_PLATFORM_AMD__)
   message(STATUS "HIP support is enabled")
-  include_directories("${ROCM_PATH}/include")
-  link_directories(
-    "${ROCM_PATH}/lib"
-  )
+
+  find_program(HIPIFY_PERL_EXECUTABLE hipify-perl)
+  if(NOT HIPIFY_PERL_EXECUTABLE)
+    message(FATAL_ERROR 
+            "hipify-perl not found.\n"
+            "Please ensure the ROCm or HIP SDK is installed and in your PATH.")
+  endif()
 endif()
 
+# This function converts given CUDA source files into HIP-compatible
+# files using hipify-perl, placing the outputs in the build directory for use in
+# project compilation. The file path changes to a new location after hipify.
+function(hipify_files input_var_name)
+    set(result_files)
+
+    foreach(input_file IN LISTS ${input_var_name})
+        file(RELATIVE_PATH rel_path ${CMAKE_SOURCE_DIR} ${input_file})
+        set(output_file "${CMAKE_BINARY_DIR}/${rel_path}")
+
+        get_filename_component(output_dir ${output_file} DIRECTORY)
+        file(MAKE_DIRECTORY ${output_dir})
+
+        add_custom_command(
+            OUTPUT ${output_file}
+            COMMAND ${HIPIFY_PERL_EXECUTABLE} ${input_file} > ${output_file}
+            DEPENDS ${input_file}
+            COMMENT "HIPifying ${input_file} → ${output_file}"
+        )
+
+        list(APPEND result_files ${output_file})
+    endforeach()
+
+    set(${input_var_name} ${result_files} PARENT_SCOPE)
+endfunction()
+
 if (USE_CXL)
   add_compile_definitions(USE_CXL)
   message(STATUS "CXL support is enabled")
diff --git a/mooncake-p2p-store/build.sh b/mooncake-p2p-store/build.sh
index 90613962..6058e3a6 100644
--- a/mooncake-p2p-store/build.sh
+++ b/mooncake-p2p-store/build.sh
@@ -39,6 +39,10 @@ if [ -d "/usr/local/cuda/lib64" ]; then
     EXT_LDFLAGS+=" -L/usr/local/cuda/lib64 -lcudart"
 fi
 
+if [ -d "/opt/rocm/lib" ]; then
+    EXT_LDFLAGS+=" -L/opt/rocm/lib64 -L/opt/rocm/lib -lamdhip64"
+fi
+
 if [ -d "/usr/local/musa/lib" ]; then
     EXT_LDFLAGS+=" -L/usr/local/musa/lib -lmusart"
 fi
diff --git a/mooncake-transfer-engine/example/CMakeLists.txt b/mooncake-transfer-engine/example/CMakeLists.txt
index 08545a37..720de438 100644
--- a/mooncake-transfer-engine/example/CMakeLists.txt
+++ b/mooncake-transfer-engine/example/CMakeLists.txt
@@ -1,29 +1,39 @@
-add_executable(transfer_engine_bench transfer_engine_bench.cpp)
+set(WORKSPACE "${CMAKE_CURRENT_SOURCE_DIR}")
+
+if (USE_HIP)
+  file(GLOB EXAMPLE_SOURCES "*.cpp")
+  hipify_files(EXAMPLE_SOURCES)
+
+  file(RELATIVE_PATH EXAMPLE_REL_PATH "${CMAKE_SOURCE_DIR}" "${CMAKE_CURRENT_SOURCE_DIR}")
+  set(WORKSPACE "${CMAKE_BINARY_DIR}/${EXAMPLE_REL_PATH}")
+endif()
+
+add_executable(transfer_engine_bench ${WORKSPACE}/transfer_engine_bench.cpp)
 target_link_libraries(transfer_engine_bench PUBLIC transfer_engine)
 
-add_executable(transfer_engine_validator transfer_engine_validator.cpp)
+add_executable(transfer_engine_validator ${WORKSPACE}/transfer_engine_validator.cpp)
 target_link_libraries(transfer_engine_validator PUBLIC transfer_engine)
 
-add_executable(transfer_engine_bench_with_notify transfer_engine_bench_with_notify.cpp)
+add_executable(transfer_engine_bench_with_notify ${WORKSPACE}/transfer_engine_bench_with_notify.cpp)
 target_link_libraries(transfer_engine_bench_with_notify PUBLIC transfer_engine)
 
-add_executable(memory_pool memory_pool.cpp)
+add_executable(memory_pool ${WORKSPACE}/memory_pool.cpp)
 target_link_libraries(memory_pool PUBLIC transfer_engine)
 
 if (USE_ASCEND)
-    add_executable(transfer_engine_ascend_one_sided transfer_engine_ascend_one_sided.cpp)
+    add_executable(transfer_engine_ascend_one_sided ${WORKSPACE}/transfer_engine_ascend_one_sided.cpp)
     target_link_libraries(transfer_engine_ascend_one_sided PUBLIC transfer_engine)
 
-    add_executable(transfer_engine_ascend_perf transfer_engine_ascend_perf.cpp)
+    add_executable(transfer_engine_ascend_perf ${WORKSPACE}/transfer_engine_ascend_perf.cpp)
     target_link_libraries(transfer_engine_ascend_perf PUBLIC transfer_engine)
 endif()
 
 if (USE_ASCEND_DIRECT)
-    add_executable(transfer_engine_ascend_direct_perf transfer_engine_ascend_direct_perf.cpp)
+    add_executable(transfer_engine_ascend_direct_perf ${WORKSPACE}/transfer_engine_ascend_direct_perf.cpp)
     target_link_libraries(transfer_engine_ascend_direct_perf PUBLIC ascendcl transfer_engine)
 endif()
 
 if (USE_ASCEND_HETEROGENEOUS)
-    add_executable(transfer_engine_heterogeneous_ascend_perf_initiator transfer_engine_heterogeneous_ascend_perf_initiator.cpp)
+    add_executable(transfer_engine_heterogeneous_ascend_perf_initiator ${WORKSPACE}/transfer_engine_heterogeneous_ascend_perf_initiator.cpp)
     target_link_libraries(transfer_engine_heterogeneous_ascend_perf_initiator PUBLIC transfer_engine)
 endif()
diff --git a/mooncake-transfer-engine/example/transfer_engine_bench.cpp b/mooncake-transfer-engine/example/transfer_engine_bench.cpp
index 21687a06..94cd1bd0 100644
--- a/mooncake-transfer-engine/example/transfer_engine_bench.cpp
+++ b/mooncake-transfer-engine/example/transfer_engine_bench.cpp
@@ -35,15 +35,15 @@
 #ifdef USE_NVMEOF
 #include <cufile.h>
 #endif
-
-#ifdef USE_MNNVL
-#include <transport/nvlink_transport/nvlink_transport.h>
-#endif
 #endif
 
 #if defined(USE_CUDA) || defined(USE_MUSA) || defined(USE_HIP)
 #include <cassert>
 
+#ifdef USE_MNNVL
+#include <transport/nvlink_transport/nvlink_transport.h>
+#endif
+
 static void checkCudaError(cudaError_t result, const char *message) {
     if (result != cudaSuccess) {
         LOG(ERROR) << message << " (Error code: " << result << " - "
diff --git a/mooncake-transfer-engine/example/transfer_engine_bench_with_notify.cpp b/mooncake-transfer-engine/example/transfer_engine_bench_with_notify.cpp
index 1310aa4d..8a77bb13 100644
--- a/mooncake-transfer-engine/example/transfer_engine_bench_with_notify.cpp
+++ b/mooncake-transfer-engine/example/transfer_engine_bench_with_notify.cpp
@@ -35,15 +35,15 @@
 #ifdef USE_NVMEOF
 #include <cufile.h>
 #endif
-
-#ifdef USE_NVLINK
-#include <transport/nvlink_transport/nvlink_transport.h>
-#endif
 #endif
 
 #if defined(USE_CUDA) || defined(USE_MUSA) || defined(USE_HIP)
 #include <cassert>
 
+#ifdef USE_MNNVL
+#include <transport/nvlink_transport/nvlink_transport.h>
+#endif
+
 static void checkCudaError(cudaError_t result, const char *message) {
     if (result != cudaSuccess) {
         LOG(ERROR) << message << " (Error code: " << result << " - "
@@ -99,7 +99,7 @@ static void *allocateMemoryPool(size_t size, int socket_id,
         int gpu_id = FLAGS_gpu_id;
         void *d_buf;
         checkCudaError(cudaSetDevice(gpu_id), "Failed to set device");
-#ifdef USE_NVLINK
+#ifdef USE_MNNVL
         d_buf = mooncake::NvlinkTransport::allocatePinnedLocalMemory(size);
 #else
         checkCudaError(cudaMalloc(&d_buf, size),
@@ -113,7 +113,7 @@ static void *allocateMemoryPool(size_t size, int socket_id,
 
 static void freeMemoryPool(void *addr, size_t size) {
 #if defined(USE_CUDA) || defined(USE_MUSA) || defined(USE_HIP)
-#ifdef USE_NVLINK
+#ifdef USE_MNNVL
     CUmemGenericAllocationHandle handle;
     auto result = cuMemRetainAllocationHandle(&handle, addr);
     if (result == CUDA_SUCCESS) {
@@ -440,7 +440,7 @@ int target() {
     }
     for (int i = 0; i < buffer_num; ++i) {
         engine->unregisterLocalMemory(addr[i]);
-#ifdef USE_NVLINK
+#ifdef USE_MNNVL
         mooncake::NvlinkTransport::freePinnedLocalMemory(addr[i]);
 #else
         freeMemoryPool(addr[i], FLAGS_buffer_size);
diff --git a/mooncake-transfer-engine/example/transfer_engine_validator.cpp b/mooncake-transfer-engine/example/transfer_engine_validator.cpp
index 3b2ee966..fc76d547 100644
--- a/mooncake-transfer-engine/example/transfer_engine_validator.cpp
+++ b/mooncake-transfer-engine/example/transfer_engine_validator.cpp
@@ -35,15 +35,14 @@
 #ifdef USE_NVMEOF
 #include <cufile.h>
 #endif
+#endif
+
+#if defined(USE_CUDA) || defined(USE_MUSA) || defined(USE_HIP)
+#include <cassert>
 
 #ifdef USE_MNNVL
 #include <transport/nvlink_transport/nvlink_transport.h>
 #endif
-#endif
-
-#if defined(USE_CUDA) || defined(USE_MUSA) || defined(USE_HIP)
-
-#include <cassert>
 
 static void checkCudaError(cudaError_t result, const char *message) {
     if (result != cudaSuccess) {
diff --git a/mooncake-transfer-engine/include/cuda_alike.h b/mooncake-transfer-engine/include/cuda_alike.h
index 3adf8a2a..6852f0da 100644
--- a/mooncake-transfer-engine/include/cuda_alike.h
+++ b/mooncake-transfer-engine/include/cuda_alike.h
@@ -11,5 +11,6 @@
 #endif
 
 #if !defined(USE_HIP) && !defined(USE_MUSA)
+#include <string>
 const static std::string GPU_PREFIX = "cuda:";
 #endif
diff --git a/mooncake-transfer-engine/include/gpu_vendor/hip.h b/mooncake-transfer-engine/include/gpu_vendor/hip.h
index 769847b1..fb67ac73 100644
--- a/mooncake-transfer-engine/include/gpu_vendor/hip.h
+++ b/mooncake-transfer-engine/include/gpu_vendor/hip.h
@@ -1,70 +1,22 @@
 #include <hip/hip_runtime.h>
+#include <string>
 
 const static std::string GPU_PREFIX = "hip:";
 
-#define CU_MEM_ACCESS_FLAGS_PROT_READWRITE hipMemAccessFlagsProtReadWrite
-#define CU_MEM_ALLOCATION_TYPE_PINNED hipMemAllocationTypePinned
-#define CU_MEM_LOCATION_TYPE_DEVICE hipMemLocationTypeDevice
-#define CU_MEM_RANGE_HANDLE_TYPE_DMA_BUF_FD hipMemRangeHandleTypeDmaBufFd
-#define CU_MEMORYTYPE_DEVICE hipMemoryTypeDevice
-#define CU_MEMORYTYPE_HOST hipMemoryTypeHost
-#define CU_POINTER_ATTRIBUTE_MEMORY_TYPE HIP_POINTER_ATTRIBUTE_MEMORY_TYPE
-#define CU_POINTER_ATTRIBUTE_RANGE_SIZE HIP_POINTER_ATTRIBUTE_RANGE_SIZE
+// hipify-perl warning: unsupported HIP identifier: cudaMemoryTypeUnregistered
+#define cudaMemoryTypeUnregistered hipMemoryTypeUnregistered
 
-#define CUdevice hipDevice_t
-#define CUdeviceptr hipDeviceptr_t
-#define CUmemAccessDesc hipMemAccessDesc
-#define CUmemAllocationProp hipMemAllocationProp
-#define CUmemGenericAllocationHandle hipMemGenericAllocationHandle_t
-#define CUmemorytype hipMemoryType
-#define CUresult hipError_t
-#define cuDeviceGet hipDeviceGet
-#define cuDeviceGetAttribute hipDeviceGetAttribute
-#define cuGetErrorString hipDrvGetErrorString
-#define cuMemAddressFree hipMemAddressFree
-#define cuMemAddressReserve hipMemAddressReserve
-#define cuMemCreate hipMemCreate
-#define cuMemGetAllocationGranularity hipMemGetAllocationGranularity
-#define cuMemGetHandleForAddressRange hipMemGetHandleForAddressRange
-#define cuMemMap hipMemMap
-#define cuMemRelease hipMemRelease
-#define cuMemSetAccess hipMemSetAccess
-#define cuMemUnmap hipMemUnmap
-#define cuPointerGetAttribute hipPointerGetAttribute
+// hipify-perl warning: unsupported HIP identifier: CU_MEM_HANDLE_TYPE_FABRIC
+// Note: HIP does not currently support multi-node (fabric-based) transfers.
+// Use POSIX file descriptor handle type for now, which only supports
+// intra-node (same-machine) memory sharing between processes.
+// TODO: Change to appropriate handle type when HIP adds multi-node support.
+#define CU_MEM_HANDLE_TYPE_FABRIC hipMemHandleTypePosixFileDescriptor
 
-#define CUDA_SUCCESS hipSuccess
-#define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer
-#define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
-#define cudaDeviceGetPCIBusId hipDeviceGetPCIBusId
-#define cudaError_t hipError_t
-#define cudaFree hipFree
-#define cudaFreeHost hipHostFree
-#define cudaGetDevice hipGetDevice
-#define cudaGetDeviceCount hipGetDeviceCount
-#define cudaGetErrorString hipGetErrorString
-#define cudaGetLastError hipGetLastError
-#define cudaHostRegister hipHostRegister
-#define cudaHostRegisterPortable hipHostRegisterPortable
-#define cudaHostUnregister hipHostUnregister
-#define cudaMalloc hipMalloc
-#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
-#define cudaMemcpy hipMemcpy
-#define cudaMemcpyAsync hipMemcpyAsync
-#define cudaMemcpyDefault hipMemcpyDefault
-#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost
-#define cudaMemcpyHostToDevice hipMemcpyHostToDevice
-#define cudaMemset hipMemset
-#define cudaMemsetAsync hipMemsetAsync
-#define cudaMemoryTypeDevice hipMemoryTypeDevice
-#define cudaMemoryTypeHost hipMemoryTypeHost
-// cudaMemoryTypeUnregistered is currently not supported as hipMemoryType enum,
-// due to HIP functionality backward compatibility.
-#define cudaMemoryTypeUnregistered 99
-#define cudaPointerAttributes hipPointerAttribute_t
-#define cudaPointerGetAttributes hipPointerGetAttributes
-#define cudaSetDevice hipSetDevice
-#define cudaStreamCreate hipStreamCreate
-#define cudaStreamDestroy hipStreamDestroy
-#define cudaStreamSynchronize hipStreamSynchronize
-#define cudaStream_t hipStream_t
-#define cudaSuccess hipSuccess
+// hipify-perl warning: unsupported HIP identifier:
+// CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_WITH_CUDA_VMM_SUPPORTED
+#define CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_WITH_CUDA_VMM_SUPPORTED \
+    hipDeviceAttributeVirtualMemoryManagementSupported
+
+// hipify-perl warning: unsupported HIP identifier: CUmemFabricHandle
+#define CUmemFabricHandle void*
diff --git a/mooncake-transfer-engine/include/transport/nvlink_transport/nvlink_transport.h b/mooncake-transfer-engine/include/transport/nvlink_transport/nvlink_transport.h
index eddac0ac..2f7d80cf 100644
--- a/mooncake-transfer-engine/include/transport/nvlink_transport/nvlink_transport.h
+++ b/mooncake-transfer-engine/include/transport/nvlink_transport/nvlink_transport.h
@@ -3,7 +3,7 @@
 #ifndef NVLINK_TRANSPORT_H_
 #define NVLINK_TRANSPORT_H_
 
-#include <cuda_runtime.h>
+#include "cuda_alike.h"
 
 #include <functional>
 #include <iostream>
diff --git a/mooncake-transfer-engine/nvlink-allocator/CMakeLists.txt b/mooncake-transfer-engine/nvlink-allocator/CMakeLists.txt
index f572fd7c..b60292b7 100644
--- a/mooncake-transfer-engine/nvlink-allocator/CMakeLists.txt
+++ b/mooncake-transfer-engine/nvlink-allocator/CMakeLists.txt
@@ -1,10 +1,28 @@
 # Build nvlink allocator and output to build directory
+
 add_custom_target(build_nvlink_allocator DEPENDS transfer_engine)
+
+get_target_property(INCLUDE_DIRS build_nvlink_allocator INCLUDE_DIRECTORIES)
+string(REPLACE ";" " " INCLUDE_DIRS_STR "${INCLUDE_DIRS}")
+
+if (USE_CUDA)
 add_custom_command(
     TARGET build_nvlink_allocator
     COMMAND ${CMAKE_COMMAND} -E make_directory ${CMAKE_CURRENT_BINARY_DIR}
-    COMMAND bash build.sh ${CMAKE_CURRENT_BINARY_DIR}
+    COMMAND bash build.sh ${CMAKE_CURRENT_BINARY_DIR} "${INCLUDE_DIRS_STR}"
     WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
     COMMENT "Building nvlink allocator to ${CMAKE_CURRENT_BINARY_DIR}"
 )
+endif()
+
+if (USE_HIP)
+add_custom_command(
+    TARGET build_nvlink_allocator
+    COMMAND ${CMAKE_COMMAND} -E make_directory ${CMAKE_CURRENT_BINARY_DIR}
+    COMMAND bash build.sh --use-hipcc ${CMAKE_CURRENT_BINARY_DIR} "${INCLUDE_DIRS_STR}"
+    WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+    COMMENT "Building nvlink allocator to ${CMAKE_CURRENT_BINARY_DIR}"
+)
+endif()
+
 set_property(TARGET build_nvlink_allocator PROPERTY EXCLUDE_FROM_ALL FALSE)
diff --git a/mooncake-transfer-engine/nvlink-allocator/build.sh b/mooncake-transfer-engine/nvlink-allocator/build.sh
index db681067..05e53403 100755
--- a/mooncake-transfer-engine/nvlink-allocator/build.sh
+++ b/mooncake-transfer-engine/nvlink-allocator/build.sh
@@ -4,11 +4,15 @@ set -e
 
 # Check for flags
 USE_NVCC=false
+USE_HIPCC=false
 CI_BUILD=false
 
 if [[ "$1" == "--use-nvcc" ]]; then
     USE_NVCC=true
     shift
+elif [[ "$1" == "--use-hipcc" ]]; then
+    USE_HIPCC=true
+    shift
 elif [[ "$1" == "--ci-build" ]]; then
     CI_BUILD=true
     shift
@@ -17,6 +21,18 @@ fi
 # Get output directory from command line argument, default to current directory
 OUTPUT_DIR=${1:-.}
 
+# Get include directories from second argument (if provided)
+INCLUDE_LIST=""
+if [ $# -ge 2 ]; then
+    INCLUDE_LIST=${2}
+fi
+
+# Process include directories into flags
+INCLUDE_FLAGS=""
+if [ -n "$INCLUDE_LIST" ]; then
+    INCLUDE_FLAGS=$(echo "$INCLUDE_LIST" | tr ' ' '\n' | sed 's/^/-I/' | paste -sd' ' -)
+fi
+
 echo "Building nvlink allocator to: $OUTPUT_DIR"
 # Create output directory if it doesn't exist
 mkdir -p "$OUTPUT_DIR"
@@ -26,13 +42,19 @@ CPP_FILE=$(dirname $(readlink -f $0))/nvlink_allocator.cpp  # get cpp file path,
 # Choose build command based on flags
 if [ "$CI_BUILD" = true ]; then
     # CI build: use nvcc without linking cuda
-    nvcc "$CPP_FILE" -o "$OUTPUT_DIR/nvlink_allocator.so" -shared -Xcompiler -fPIC -I/usr/local/cuda/include -L/usr/local/cuda-12.8/targets/x86_64-linux/lib -lcuda
+    nvcc "$CPP_FILE" -o "$OUTPUT_DIR/nvlink_allocator.so" -shared -Xcompiler -fPIC -I/usr/local/cuda/include ${INCLUDE_FLAGS} -L/usr/local/cuda-12.8/targets/x86_64-linux/lib -lcuda -DUSE_CUDA=1
 elif [ "$USE_NVCC" = true ]; then
     # Regular nvcc build with cuda linking
-    nvcc "$CPP_FILE" -o "$OUTPUT_DIR/nvlink_allocator.so" -shared -Xcompiler -fPIC -lcuda -I/usr/local/cuda/include
+    nvcc "$CPP_FILE" -o "$OUTPUT_DIR/nvlink_allocator.so" -shared -Xcompiler -fPIC -lcuda -I/usr/local/cuda/include ${INCLUDE_FLAGS} -DUSE_CUDA=1
+elif [ "$USE_HIPCC" = true ]; then
+    hipify-perl "$CPP_FILE" > "${OUTPUT_DIR}/nvlink_allocator.cpp"
+    hipcc "$OUTPUT_DIR/nvlink_allocator.cpp" -o "$OUTPUT_DIR/nvlink_allocator.so" -shared -fPIC -lamdhip64 -I/opt/rocm/include ${INCLUDE_FLAGS} -DUSE_HIP=1
 else
     # Default g++ build
-    g++ "$CPP_FILE" -o "$OUTPUT_DIR/nvlink_allocator.so" --shared -fPIC -lcuda -I/usr/local/cuda/include
+    # Add include directory for cuda_alike.h (relative to build.sh location)
+    SCRIPT_DIR=$(dirname $(readlink -f $0))
+    DEFAULT_INCLUDE_DIR="${SCRIPT_DIR}/../include"
+    g++ "$CPP_FILE" -o "$OUTPUT_DIR/nvlink_allocator.so" --shared -fPIC -lcuda -I/usr/local/cuda/include -I${DEFAULT_INCLUDE_DIR} ${INCLUDE_FLAGS} -DUSE_CUDA=1
 fi
 
 if [ $? -eq 0 ]; then
diff --git a/mooncake-transfer-engine/nvlink-allocator/nvlink_allocator.cpp b/mooncake-transfer-engine/nvlink-allocator/nvlink_allocator.cpp
index 12698e40..474a4c23 100644
--- a/mooncake-transfer-engine/nvlink-allocator/nvlink_allocator.cpp
+++ b/mooncake-transfer-engine/nvlink-allocator/nvlink_allocator.cpp
@@ -1,5 +1,4 @@
-#include <cuda.h>
-#include <cuda_runtime_api.h>
+#include "cuda_alike.h"
 #include <sys/types.h>
 
 #include <iostream>
diff --git a/mooncake-transfer-engine/src/CMakeLists.txt b/mooncake-transfer-engine/src/CMakeLists.txt
index abd9e8b5..7a5dd1c7 100644
--- a/mooncake-transfer-engine/src/CMakeLists.txt
+++ b/mooncake-transfer-engine/src/CMakeLists.txt
@@ -4,6 +4,10 @@ add_subdirectory(transport)
 
 SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
 
+if (USE_HIP)
+  hipify_files(ENGINE_SOURCES)
+endif()
+
 add_library(transfer_engine ${ENGINE_SOURCES} $<TARGET_OBJECTS:transport>)
 if (BUILD_SHARED_LIBS)
   install(TARGETS transfer_engine DESTINATION lib)
@@ -57,8 +61,8 @@ if (USE_MUSA)
 endif()
 
 if (USE_HIP)
-  target_include_directories(transfer_engine PRIVATE "${ROCM_PATH}/include")
-  target_link_libraries(transfer_engine PUBLIC amdhip64 rt)
+  target_include_directories(transfer_engine PRIVATE ${HIP_INCLUDE_DIRS})
+  target_link_libraries(transfer_engine PUBLIC hip::host rt)
 endif()
 
 if (USE_ASCEND)
diff --git a/mooncake-transfer-engine/src/transport/nvlink_transport/CMakeLists.txt b/mooncake-transfer-engine/src/transport/nvlink_transport/CMakeLists.txt
index 799bbe29..0b9c03a3 100644
--- a/mooncake-transfer-engine/src/transport/nvlink_transport/CMakeLists.txt
+++ b/mooncake-transfer-engine/src/transport/nvlink_transport/CMakeLists.txt
@@ -1,5 +1,15 @@
 file(GLOB NVLINK_SOURCES "*.cpp")
 
+if (USE_HIP)
+  hipify_files(NVLINK_SOURCES)
+endif()
 
 add_library(nvlink_transport OBJECT ${NVLINK_SOURCES})
+
+if (USE_CUDA)
 target_include_directories(nvlink_transport PUBLIC  CUDA::cudart "/usr/local/cuda/include")
+endif()
+
+if (USE_HIP)
+target_include_directories(nvlink_transport PUBLIC ${HIP_INCLUDE_DIRS})
+endif()
diff --git a/mooncake-transfer-engine/src/transport/nvlink_transport/nvlink_transport.cpp b/mooncake-transfer-engine/src/transport/nvlink_transport/nvlink_transport.cpp
index 9d103047..b7213c06 100644
--- a/mooncake-transfer-engine/src/transport/nvlink_transport/nvlink_transport.cpp
+++ b/mooncake-transfer-engine/src/transport/nvlink_transport/nvlink_transport.cpp
@@ -15,8 +15,7 @@
 #include "transport/nvlink_transport/nvlink_transport.h"
 
 #include <bits/stdint-uintn.h>
-#include <cuda.h>
-#include <cuda_runtime.h>
+#include "cuda_alike.h"
 #include <glog/logging.h>
 
 #include <algorithm>
@@ -69,6 +68,7 @@ static bool supportFabricMem() {
         return false;
     }
 
+#ifdef USE_CUDA
     for (int device_id = 0; device_id < num_devices; ++device_id) {
         int device_support_fabric_mem = 0;
         cuDeviceGetAttribute(&device_support_fabric_mem,
@@ -78,6 +78,7 @@ static bool supportFabricMem() {
             return false;
         }
     }
+#endif
     return true;
 }
 
diff --git a/mooncake-transfer-engine/src/transport/tcp_transport/CMakeLists.txt b/mooncake-transfer-engine/src/transport/tcp_transport/CMakeLists.txt
index 23fa1b4a..f2c21ea0 100644
--- a/mooncake-transfer-engine/src/transport/tcp_transport/CMakeLists.txt
+++ b/mooncake-transfer-engine/src/transport/tcp_transport/CMakeLists.txt
@@ -1,4 +1,8 @@
 file(GLOB TCP_SOURCES "*.cpp")
 
+if (USE_HIP)
+  hipify_files(TCP_SOURCES)
+endif()
+
 add_library(tcp_transport OBJECT ${TCP_SOURCES})
 target_link_libraries(tcp_transport PRIVATE JsonCpp::JsonCpp yalantinglibs::yalantinglibs)
diff --git a/mooncake-transfer-engine/tests/CMakeLists.txt b/mooncake-transfer-engine/tests/CMakeLists.txt
index 04e47173..6594d887 100644
--- a/mooncake-transfer-engine/tests/CMakeLists.txt
+++ b/mooncake-transfer-engine/tests/CMakeLists.txt
@@ -1,55 +1,65 @@
-add_executable(rdma_transport_test rdma_transport_test.cpp)
+set(WORKSPACE "${CMAKE_CURRENT_SOURCE_DIR}")
+
+if (USE_HIP)
+  file(GLOB TEST_SOURCES "*.cpp")
+  hipify_files(TEST_SOURCES)
+
+  file(RELATIVE_PATH EXAMPLE_REL_PATH "${CMAKE_SOURCE_DIR}" "${CMAKE_CURRENT_SOURCE_DIR}")
+  set(WORKSPACE "${CMAKE_BINARY_DIR}/${EXAMPLE_REL_PATH}")
+endif()
+
+add_executable(rdma_transport_test ${WORKSPACE}/rdma_transport_test.cpp)
 target_link_libraries(rdma_transport_test PUBLIC transfer_engine)
 # add_test(NAME rdma_transport_test COMMAND rdma_transport_test)
 
-add_executable(transport_uint_test transport_uint_test.cpp)
+add_executable(transport_uint_test ${WORKSPACE}/transport_uint_test.cpp)
 target_link_libraries(transport_uint_test PUBLIC transfer_engine gtest gtest_main )
 add_test(NAME transport_uint_test COMMAND transport_uint_test)
 
-add_executable(rdma_transport_test2 rdma_transport_test2.cpp)
+add_executable(rdma_transport_test2 ${WORKSPACE}/rdma_transport_test2.cpp)
 target_link_libraries(rdma_transport_test2 PUBLIC transfer_engine gtest gtest_main )
 # add_test(NAME rdma_transport_test2 COMMAND rdma_transport_test2)
 
-add_executable(rdma_loopback_test rdma_loopback_test.cpp)
+add_executable(rdma_loopback_test ${WORKSPACE}/rdma_loopback_test.cpp)
 target_link_libraries(rdma_loopback_test PUBLIC transfer_engine gtest gtest_main )
 # add_test(NAME rdma_loopback_test COMMAND rdma_loopback_test)
 
 if (USE_CXL)
-    add_executable(cxl_transport_test cxl_transport_test.cpp)
+    add_executable(cxl_transport_test ${WORKSPACE}/cxl_transport_test.cpp)
     target_link_libraries(cxl_transport_test PUBLIC transfer_engine gtest gtest_main )
     add_test(NAME cxl_transport_test COMMAND cxl_transport_test)
 endif()
 
 if (USE_NVMEOF)
-    add_executable(nvmeof_transport_test nvmeof_transport_test.cpp)
+    add_executable(nvmeof_transport_test ${WORKSPACE}/nvmeof_transport_test.cpp)
     target_link_libraries(nvmeof_transport_test PUBLIC transfer_engine gtest gtest_main )
     # add_test(NAME nvmeof_transport_test COMMAND nvmeof_transport_test)
 endif()
 
 if (USE_TCP)
-add_executable(tcp_transport_test tcp_transport_test.cpp)
+add_executable(tcp_transport_test ${WORKSPACE}/tcp_transport_test.cpp)
 target_link_libraries(tcp_transport_test PUBLIC transfer_engine gtest gtest_main )
 add_test(NAME tcp_transport_test COMMAND tcp_transport_test)
 endif()
 
 if (USE_MNNVL)
-    add_executable(nvlink_transport_test nvlink_transport_test.cpp)
+    add_executable(nvlink_transport_test ${WORKSPACE}/nvlink_transport_test.cpp)
     target_link_libraries(nvlink_transport_test PUBLIC transfer_engine gtest gtest_main )
     add_test(NAME nvlink_transport_test COMMAND nvlink_transport_test)  
 endif()
 
-add_executable(transfer_metadata_test transfer_metadata_test.cpp)
+add_executable(transfer_metadata_test ${WORKSPACE}/transfer_metadata_test.cpp)
 target_link_libraries(transfer_metadata_test PUBLIC transfer_engine gtest gtest_main)
 add_test(NAME transfer_metadata_test COMMAND transfer_metadata_test)
 
-add_executable(topology_test topology_test.cpp)
+add_executable(topology_test ${WORKSPACE}/topology_test.cpp)
 target_link_libraries(topology_test PUBLIC transfer_engine gtest gtest_main)
 add_test(NAME topology_test COMMAND topology_test)
 
-add_executable(memory_location_test memory_location_test.cpp)
+add_executable(memory_location_test ${WORKSPACE}/memory_location_test.cpp)
 target_link_libraries(memory_location_test PUBLIC transfer_engine gtest gtest_main)
 add_test(NAME memory_location_test COMMAND memory_location_test)
 
-add_executable(common_test common_test.cpp)
+add_executable(common_test ${WORKSPACE}/common_test.cpp)
 target_link_libraries(common_test PUBLIC transfer_engine gtest gtest_main)
 add_test(NAME common_test COMMAND common_test)
diff --git a/mooncake-transfer-engine/tests/nvlink_transport_test.cpp b/mooncake-transfer-engine/tests/nvlink_transport_test.cpp
index b49341e8..bce7029b 100644
--- a/mooncake-transfer-engine/tests/nvlink_transport_test.cpp
+++ b/mooncake-transfer-engine/tests/nvlink_transport_test.cpp
@@ -1,11 +1,11 @@
 #include <gflags/gflags.h>
 #include <glog/logging.h>
 #include <gtest/gtest.h>
-#include <cuda_runtime.h>
 #include <thread>
 #include <memory>
 #include <cstring>
 
+#include "cuda_alike.h"
 #include "transfer_engine.h"
 #include "transport/transport.h"