diff --git a/docs/build.md b/docs/build.md
index 3c5c17a89..17f645239 100644
--- a/docs/build.md
+++ b/docs/build.md
@@ -6,11 +6,6 @@
 - Torch 2.4.1
 - Clang 19
 
-#### if for AMD GPU:
-- ROCM 6.3.0
-- Torch 2.4.1 with ROCM support
-
-
 
 Dependencies with other versions may also work well, but this is not guaranteed. If you find any problem in installing, please tell us in Issues.
 
@@ -26,10 +21,7 @@ Dependencies with other versions may also work well, but this is not guaranteed.
     pip3 install black "clang-format==19.1.2" pre-commit ruff yapf==0.43
     pip3 install ninja cmake wheel pybind11 cuda-python==12.4 numpy chardet pytest
     ```
-    for AMD GPU, use torch with rocm support and hip-python
-    ```sh
-    python3 -m pip install -i https://test.pypi.org/simple hip-python>=6.3.0
-    ```
+
 4. Apply NVSHMEM fix
 (Disclaimer: This step is because of NVSHMEM license requirements, it is illegal to release any modified codes or patch.)
 
@@ -84,8 +76,6 @@ Dependencies with other versions may also work well, but this is not guaranteed.
     export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/llvm-project/build/lib
     ```
 
-    For ROCMSHMEM on AMD GPU, no explicit build required as the building process is integrated with Triton-distributed.
-
 6. Build Triton-distributed
     Then you can build Triton-distributed.
     ```sh
@@ -114,20 +104,13 @@ This example runs on a single node with 8 H800 GPUs.
 ```sh
 bash ./third_party/distributed/launch.sh ./third_party/distributed/distributed/test/nvidia/test_ag_gemm_intra_node.py --case correctness_tma
 ```
-For AMD CDNA3 GPUs:
-```sh
-bash ./third_party/distributed/launch_amd.sh ./third_party/distributed/distributed/test/amd/test_ag_gemm_intra_node.py 8192 53248 16384
-```
 
 #### GEMM ReduceScatter example on single node
 This example runs on a single node with 8 H800 GPUs.
 ```sh
 bash ./third_party/distributed/launch.sh ./third_party/distributed/distributed/test/nvidia/test_gemm_rs_multi_node.py 8192 8192 29568
 ```
-For AMD CDNA3 GPUs:
-```sh
-bash ./third_party/distributed/launch_amd.sh ./third_party/distributed/distributed/test/amd/test_gemm_rs_intra_node.py 8192 3584 14336
-```
+
 #### NVSHMEM example in Triton-distributed
 ```sh
 bash ./third_party/distributed/launch.sh ./third_party/distributed/distributed/test/nvidia/test_nvshmem_api.py
@@ -173,40 +156,4 @@ bash ./third_party/distributed/launch.sh ./third_party/distributed/distributed/t
 # moe rs
 bash ./third_party/distributed/launch.sh ./third_party/distributed/distributed/test/test_moe_reduce_rs_intra_node.py 8192 2048 1536 32 2
 bash ./third_party/distributed/launch.sh ./third_party/distributed/distributed/test/test_moe_reduce_rs_intra_node.py 8192 2048 1536 32 2 --check
-```
-
-## To use Triton-distributed with the AMD backend:
-- Starting from the rocm/pytorch:rocm6.1_ubuntu22.04_py3.10_pytorch_2.4 Docker container
-#### Steps:
-1. Clone the repo
-```sh
-git clone https://github.com/ByteDance-Seed/Triton-distributed.git
-```
-2. Update submodules
-```sh
-cd Triton-distributed/
-git submodule update --init --recursive
-```
-3. Install dependencies
-```sh
-sudo apt-get update -y
-sudo apt install -y libopenmpi-dev
-pip3 install --pre torch --index-url https://download.pytorch.org/whl/nightly/rocm6.3 --no-deps
-./third_party/rocshmem_bind/build.sh
-python3 -m pip install -i https://test.pypi.org/simple hip-python~=6.3.2 (or whatever Rocm version you have)
-pip3 install pybind11
-```
-4. Build Triton-distributed
-```sh
-pip3 install -e python --verbose --no-build-isolation
-```
-### Test your installation
-#### GEMM ReduceScatter example on single node
-```sh
-bash ./third_party/distributed/launch_amd.sh ./third_party/distributed/distributed/test/amd/test_ag_gemm_intra_node.py 8192 8192 29568
- ```
-and see the following (reduced) output
-```sh
-torchrun --node_rank=0 --nproc_per_node=8 --nnodes=1 ./third_party/distributed/distributed/test/amd/test_ag_gemm_intra_node.py 8192 8192 29568
-✅ Triton and Torch match
-```
+```
\ No newline at end of file
diff --git a/docs/build_amd.md b/docs/build_amd.md
new file mode 100644
index 000000000..a11d21249
--- /dev/null
+++ b/docs/build_amd.md
@@ -0,0 +1,54 @@
+# Build Triton-distributed (RocSHMEM)
+
+## The best practice to use Trition-distributed in AMD GPU
+
+- ROCm 6.3.3
+- torch-2.6/2.8 (torch-2.6 has major improvement, compatible with SGLang, vLLM)
+- python3.12.8
+- MI300X/MI325X
+
+Dependencies with other versions may also work well, but this is not guaranteed. If you find any problem in installing, please tell us in Issues.
+
+## Setup without docker
+
+1. make sure torch-rocm is installed for ROCm SDK 6.3.3
+    ```sh
+    pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.3
+    ```
+2. install ompi for ROCm SDK
+    ```
+    sudo apt-get update -y && \
+    sudo apt install -y libopenmpi-dev
+    ```
+3. install other dependencies
+   ```
+   python3 -m pip install -i https://test.pypi.org/simple hip-python~=6.3.3 # or whatever Rocm version you have
+   pip3 install pybind11
+   ```
+
+#### Warnning of install inside existing dockder
+
+Make sure following repositories granted permission to clone submodules
+
+```
+export TRITON_DIST_HOME=$(readlink -f `pwd`)
+
+git config --global --add safe.directory $TRITON_DIST_HOME/Triton-distributed/3rdparty/rocshmem
+git config --global --add safe.directory $TRITON_DIST_HOME/Triton-distributed/3rdparty/triton
+git config --global --add safe.directory $TRITON_DIST_HOME/Triton-distributed
+
+git submodule update --init --recursive
+```
+
+## Build
+
+> python3 python/setup.py build_ext
+
+## Test
+
+Currently only single node build supported, multi-node build will be supported soon.
+
+- Single node test
+  ```
+  bash ./scripts/launch_amd.sh python/triton_dist/test/amd/test_ag_gemm_intra_node.py 8192 8192 29568
+  ```
\ No newline at end of file
diff --git a/python/build_helpers.py b/python/build_helpers.py
index 8b50d735c..160b757a3 100644
--- a/python/build_helpers.py
+++ b/python/build_helpers.py
@@ -46,5 +46,8 @@ def copy_apply_patches():
         for file in files:
             source_file = os.path.join(root, file)
             target_file = os.path.join(target_dir, file)
-            shutil.copy2(source_file, target_file)
+            try:
+                shutil.copy2(source_file, target_file)
+            except Exception:
+                shutil.copyfile(source_file, target_file)
             print(f"Copied {source_file} to {target_file}")
diff --git a/shmem/rocshmem_bind/scripts/build_rshm_ipc_single.sh b/shmem/rocshmem_bind/scripts/build_rshm_ipc_single.sh
index 515efae72..963013b0e 100755
--- a/shmem/rocshmem_bind/scripts/build_rshm_ipc_single.sh
+++ b/shmem/rocshmem_bind/scripts/build_rshm_ipc_single.sh
@@ -9,6 +9,8 @@ else
   install_path=$1
 fi
 
+hip_cmake_path="/opt/rocm/lib/cmake/hip;/opt/rocm/lib/cmake/rocprim;/opt/rocm/lib/cmake/rocthrust"
+
 src_path=$(dirname "$(realpath $0)")/../../../3rdparty/rocshmem/
 
 cmake \
@@ -29,6 +31,7 @@ cmake \
     -DUSE_SINGLE_NODE=ON \
     -DUSE_HOST_SIDE_HDP_FLUSH=OFF \
     -DBUILD_LOCAL_GPU_TARGET_ONLY=ON \
+    -DCMAKE_PREFIX_PATH="$hip_cmake_path" \
     $src_path
 cmake --build . --parallel
 cmake --install .