diff --git a/.github/workflows/publish.yml b/.github/workflows/publish.yml
new file mode 100644
index 00000000..22720a5c
--- /dev/null
+++ b/.github/workflows/publish.yml
@@ -0,0 +1,34 @@
+name: Publish Python Package
+
+on:
+  release:
+    types: [created]
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Check out repository
+        uses: actions/checkout@v3
+
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: '3.10'  # Specify the Python version you need
+
+      - name: Install dependencies
+        run: |
+          python3 -m pip install --upgrade pip
+          python3 -m pip install -r requirements.txt
+
+      - name: Build package
+        run: python3 -m build -n --sdist
+
+      - name: Publish to PyPI
+        env:
+          TWINE_USERNAME: __token__
+          TWINE_PASSWORD: ${{ secrets.PIP_TOKEN }}
+        run: |
+          python3 -m pip install twine
+          twine upload dist/*
\ No newline at end of file
diff --git a/docs/source/install.rst b/docs/source/install.rst
index 0b7867d5..cd8ba8ee 100644
--- a/docs/source/install.rst
+++ b/docs/source/install.rst
@@ -42,7 +42,22 @@ If you prefer not to use Docker, you can also install ReaL from PyPI or from the
 
    We don't upload a pre-built wheel to PyPI, so the installation will require compiling the C++ and CUDA extensions. If CUDA is not available on your machine, only the C++ extension will be installed.
 
-Install from PyPI:
+First, clone the repository and install all dependencies:
+
+.. code-block:: console
+
+   $ git clone https://github.com/openpsi-project/ReaLHF
+   $ cd ReaLHF
+   $ python3 -m pip install -r requirements.txt
+
+On a GPU machine, also install the requirement CUDA runtime packages:
+
+.. code-block:: console
+
+   $ python3 -m pip install git+https://github.com/NVIDIA/TransformerEngine.git@v1.7 --no-deps
+   $ python3 -m pip install flash_attn --no-build-isolation
+
+Install ReaLHF from PyPI:
 
 .. code-block:: console
 
diff --git a/pyproject.toml b/pyproject.toml
index 289954c8..cb7ddb38 100755
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -6,7 +6,7 @@ build-backend = "setuptools.build_meta"
 name = "realhf"
 description = "ReaL: Efficient RLHF Training of Large Language Models with Parameter Reallocation"
 readme = "README.md"
-requires-python = ">=3.10"
+requires-python = ">=3.10,<3.12"
 dynamic = ["version"]
 authors = [
     { name = "Zhiyu Mei", email = "meizy20@mails.tsinghua.edu.cn" },
diff --git a/realhf/__init__.py b/realhf/__init__.py
index d57fd414..f5e38bbc 100644
--- a/realhf/__init__.py
+++ b/realhf/__init__.py
@@ -14,11 +14,11 @@
     OptimizerConfig,
     ParallelismConfig,
 )
+from .base.namedarray import NamedArray
 from .experiments.common.common import CommonExperimentConfig
 from .experiments.common.dpo_exp import DPOConfig
 from .experiments.common.ppo_exp import PPOConfig, PPOHyperparameters
 from .experiments.common.rw_exp import RWConfig
 from .experiments.common.sft_exp import SFTConfig
-from .base.namedarray import NamedArray
 
 __version__ = "0.1.0"
diff --git a/setup.py b/setup.py
index 340f6057..936364b5 100644
--- a/setup.py
+++ b/setup.py
@@ -212,9 +212,6 @@ def get_torch_arch_list() -> Set[str]:
         with contextlib.suppress(ValueError):
             torch_cpp_ext.COMMON_NVCC_FLAGS.remove(flag)
 
-with open("requirements.txt", "r") as f:
-    dependencies = f.read().splitlines()
-
 os.makedirs(os.path.join(ROOT_DIR, "realhf", "_C"), exist_ok=True)
 if _is_cuda():
     cr_extension = CUDAExtension(
@@ -291,5 +288,4 @@ def get_torch_arch_list() -> Set[str]:
     ext_modules=ext_modules,
     cmdclass={"build_ext": BuildExtension},
     packages=setuptools.find_packages(),
-    install_requires=dependencies,
 )