diff --git a/docs/quickstart.md b/docs/quickstart.md
index c9c98128..83a08d6a 100644
--- a/docs/quickstart.md
+++ b/docs/quickstart.md
@@ -25,9 +25,9 @@
         ```bash
         sudo apt-get install libnuma-dev
         ```
-    * (Optional, for [building the Python module](#install-from-source-python-module)) Python >= 3.8 and Python Development Package
+    * (Optional, for [building the Python module](#install-from-source-python-module)) Python >= 3.10 and Python Development Package
         ```bash
-        sudo apt-get satisfy "python3 (>=3.8), python3-dev (>=3.8)"
+        sudo apt-get satisfy "python3 (>=3.10), python3-dev (>=3.10)"
         ```
         If you don't want to build Python module, you need to set `-DMSCCLPP_BUILD_PYTHON_BINDINGS=OFF` in your `cmake` command (see details in [Install from Source](#install-from-source)).
     * (Optional, for benchmarks) MPI
@@ -100,13 +100,30 @@ There are a few optional CMake options you can set:
 (install-from-source-python-module)=
 ## Install from Source (Python Module)
 
-Python 3.8 or later is required.
+Python 3.10 or later is required.
 
 ```bash
-# For NVIDIA platforms
-$ python -m pip install .
-# For AMD platforms, set the C++ compiler to HIPCC
-$ CXX=/opt/rocm/bin/hipcc python -m pip install .
+# For NVIDIA platforms (specify your CUDA version)
+$ python -m pip install ".[cuda12]"
+# For AMD platforms
+$ CXX=/opt/rocm/bin/hipcc python -m pip install ".[rocm6]"
+```
+
+> **Note:** A platform extra (`cuda11`, `cuda12`, `cuda13`, or `rocm6`) is required to install CuPy.
+> The CUDA extras install pre-built CuPy wheels. The `rocm6` extra installs CuPy from source,
+> which requires ROCm and may take longer. Running `pip install .` without an extra will not install CuPy.
+
+Optional extras can be installed by specifying them in brackets. Available extras:
+- **`cuda11`**, **`cuda12`**, **`cuda13`**: Install a pre-built CuPy package for your CUDA version.
+- **`rocm6`**: Install CuPy from source for AMD ROCm platforms.
+- **`benchmark`**: Install benchmark dependencies (mpi4py, prettytable, netifaces, matplotlib).
+- **`test`**: Install test dependencies (pytest, mpi4py, netifaces).
+
+```bash
+# Example: install with CUDA 12 and benchmark extras
+$ python -m pip install ".[cuda12,benchmark]"
+# Example: install with all extras for testing on CUDA 12
+$ python -m pip install ".[cuda12,benchmark,test]"
 ```
 
 (vscode-dev-container)=
@@ -158,8 +175,9 @@ $ mpirun -np 16 -npernode 8 -hostfile hostfile ./bin/mp_unit_tests -ip_port 10.0
 [Install the MSCCL++ Python package](#install-from-source-python-module) and run our Python AllReduce benchmark as follows. It requires MPI on the system.
 
 ```bash
-# Choose `requirements_*.txt` according to your CUDA/ROCm version.
-$ python3 -m pip install -r ./python/requirements_cuda12.txt
+# Install with benchmark dependencies and the appropriate CUDA/ROCm extras.
+# Replace `cuda12` with your platform: cuda11, cuda12, cuda13, or rocm6.
+$ python3 -m pip install ".[cuda12,benchmark,test]"
 $ mpirun -tag-output -np 8 python3 ./python/mscclpp_benchmark/allreduce_bench.py
 ```
 
diff --git a/pyproject.toml b/pyproject.toml
index 651fec3b..0ea569cb 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -12,7 +12,30 @@ build-backend = "scikit_build_core.build"
 name = "mscclpp"
 dynamic = ["version"]
 description = "MSCCL++ Python API"
-requires-python = ">=3.8"
+requires-python = ">=3.10"
+dependencies = [
+    "numpy",
+    "blake3",
+    "pybind11",
+    "sortedcontainers",
+]
+
+[project.optional-dependencies]
+cuda11 = ["cupy-cuda11x"]
+cuda12 = ["cupy-cuda12x"]
+cuda13 = ["cupy-cuda13x"]
+rocm6 = ["cupy"]
+benchmark = [
+    "mpi4py",
+    "prettytable",
+    "netifaces",
+    "matplotlib",
+]
+test = [
+    "pytest",
+    "mpi4py",
+    "netifaces",
+]
 
 [tool.setuptools_scm]
 write_to = "python/mscclpp/_version.py"
@@ -40,5 +63,5 @@ MSCCLPP_BUILD_TESTS = "OFF"
 
 [tool.black]
 line-length = 120
-target-version = ['py38']
+target-version = ['py310']
 include = '\.pyi?$'
diff --git a/python/csrc/CMakeLists.txt b/python/csrc/CMakeLists.txt
index 44fb150f..7c7bf3b9 100644
--- a/python/csrc/CMakeLists.txt
+++ b/python/csrc/CMakeLists.txt
@@ -1,7 +1,7 @@
 # Copyright (c) Microsoft Corporation.
 # Licensed under the MIT license.
 
-find_package(Python 3.8 COMPONENTS Interpreter Development.Module REQUIRED)
+find_package(Python 3.10 COMPONENTS Interpreter Development.Module REQUIRED)
 include(FetchContent)
 FetchContent_Declare(nanobind GIT_REPOSITORY https://github.com/wjakob/nanobind.git GIT_TAG v1.9.2)
 FetchContent_MakeAvailable(nanobind)
diff --git a/python/mscclpp/_core/compiler.py b/python/mscclpp/_core/compiler.py
index b2da976d..3b77ce8e 100644
--- a/python/mscclpp/_core/compiler.py
+++ b/python/mscclpp/_core/compiler.py
@@ -192,6 +192,9 @@ class NativeCodeCompiler:
     """
 
     def __init__(self):
+        self._initialized = False
+
+    def _do_init(self):
         self._is_hip = cp.cuda.runtime.is_hip
         self._device_arch = get_device_arch()
         self._compiler = self._get_compiler()
@@ -226,6 +229,7 @@ class NativeCodeCompiler:
         ]
         self._cache_dir = Path(env().cache_dir) / "native"
         self._cache_dir.mkdir(parents=True, exist_ok=True)
+        self._initialized = True
 
     def _get_compiler(self) -> str:
         """Get the path to the appropriate compiler.
@@ -246,6 +250,8 @@ class NativeCodeCompiler:
         Returns:
             str: The GPU architecture string (e.g., "sm_90" for NVIDIA or "gfx90a" for AMD).
         """
+        if not self._initialized:
+            self._do_init()
         return self._device_arch
 
     def __call__(self, name: str, file: str, **kwds):
@@ -290,6 +296,8 @@ class NativeCodeCompiler:
             >>> # Use the module to create an algorithm
             >>> algo = module.create_allreduce_algorithm(comm, buffer, size)
         """
+        if not self._initialized:
+            self._do_init()
         if not os.path.isfile(file):
             raise FileNotFoundError(f"The specified source file does not exist: {file}")
 
diff --git a/python/requirements_cuda11.txt b/python/requirements_cuda11.txt
index 4e2e9371..a9786071 100644
--- a/python/requirements_cuda11.txt
+++ b/python/requirements_cuda11.txt
@@ -5,6 +5,6 @@ netifaces
 pytest
 numpy
 matplotlib
-sortedcontainers @ git+https://github.com/grantjenks/python-sortedcontainers.git@3ac358631f58c1347f1d6d2d92784117db0f38ed
+sortedcontainers
 blake3
 pybind11
\ No newline at end of file
diff --git a/python/requirements_cuda12.txt b/python/requirements_cuda12.txt
index e1c9b726..71572714 100644
--- a/python/requirements_cuda12.txt
+++ b/python/requirements_cuda12.txt
@@ -5,6 +5,6 @@ netifaces
 pytest
 numpy
 matplotlib
-sortedcontainers @ git+https://github.com/grantjenks/python-sortedcontainers.git@3ac358631f58c1347f1d6d2d92784117db0f38ed
+sortedcontainers
 blake3
 pybind11
\ No newline at end of file
diff --git a/python/requirements_cuda13.txt b/python/requirements_cuda13.txt
index 49cf13bc..95e99533 100644
--- a/python/requirements_cuda13.txt
+++ b/python/requirements_cuda13.txt
@@ -5,6 +5,6 @@ netifaces
 pytest
 numpy
 matplotlib
-sortedcontainers @ git+https://github.com/grantjenks/python-sortedcontainers.git@3ac358631f58c1347f1d6d2d92784117db0f38ed
+sortedcontainers
 blake3
 pybind11
\ No newline at end of file
diff --git a/python/requirements_rocm6.txt b/python/requirements_rocm6.txt
index 7ed4fef3..757d4e26 100644
--- a/python/requirements_rocm6.txt
+++ b/python/requirements_rocm6.txt
@@ -5,6 +5,6 @@ netifaces
 pytest
 numpy
 matplotlib
-sortedcontainers @ git+https://github.com/grantjenks/python-sortedcontainers.git@3ac358631f58c1347f1d6d2d92784117db0f38ed
+sortedcontainers
 blake3
 pybind11
\ No newline at end of file
diff --git a/python/test/CMakeLists.txt b/python/test/CMakeLists.txt
index be62aea9..e55711d2 100644
--- a/python/test/CMakeLists.txt
+++ b/python/test/CMakeLists.txt
@@ -1,7 +1,7 @@
 # Copyright (c) Microsoft Corporation.
 # Licensed under the MIT license.
 
-find_package(Python 3.8 COMPONENTS Interpreter Development.Module REQUIRED)
+find_package(Python 3.10 COMPONENTS Interpreter Development.Module REQUIRED)
 include(FetchContent)
 FetchContent_Declare(nanobind GIT_REPOSITORY https://github.com/wjakob/nanobind.git GIT_TAG v1.4.0)
 FetchContent_MakeAvailable(nanobind)
diff --git a/test/deploy/setup.sh b/test/deploy/setup.sh
index bc29efd8..2a88a310 100644
--- a/test/deploy/setup.sh
+++ b/test/deploy/setup.sh
@@ -50,12 +50,6 @@ elif [ ${PEER_ACCESS_EXIT_CODE} -ne 0 ]; then
 fi
 make -C /root/mscclpp/tools/peer-access-test clean
 
-if [[ "${CUDA_VERSION}" == *"11."* ]]; then
-    pip3 install -r /root/mscclpp/python/requirements_cuda11.txt
-elif [[ "${CUDA_VERSION}" == *"12."* ]]; then
-    pip3 install -r /root/mscclpp/python/requirements_cuda12.txt
-fi
-
 if [ "${PLATFORM}" == "rocm" ]; then
     export CXX=/opt/rocm/bin/hipcc
 fi
@@ -65,7 +59,19 @@ if [ -f "${PIP_CMAKE_ARGS_FILE}" ]; then
     export CMAKE_ARGS="$(cat ${PIP_CMAKE_ARGS_FILE})"
     echo "Using CMAKE_ARGS: ${CMAKE_ARGS}"
 fi
-cd /root/mscclpp && pip3 install .
+
+cd /root/mscclpp
+if [[ "${CUDA_VERSION}" == *"11."* ]]; then
+    pip3 install ".[cuda11,benchmark,test]"
+elif [[ "${CUDA_VERSION}" == *"12."* ]]; then
+    pip3 install ".[cuda12,benchmark,test]"
+elif [[ "${CUDA_VERSION}" == *"13."* ]]; then
+    pip3 install ".[cuda13,benchmark,test]"
+elif [ "${PLATFORM}" == "rocm" ]; then
+    pip3 install ".[rocm6,benchmark,test]"
+else
+    pip3 install ".[benchmark,test]"
+fi
 pip3 install setuptools_scm
 python3 -m setuptools_scm --force-write-version-files