Add documentation on how to do incremental builds

docs/source/getting_started/installation.rst

@@ -67,3 +67,13 @@ You can also build and install vLLM from source:
 
         $ # Use `--ipc=host` to make sure the shared memory is large enough.
         $ docker run --gpus all -it --rm --ipc=host nvcr.io/nvidia/pytorch:23.10-py3
+
+.. note::
+    If you are developing the C++ backend of vLLM, consider building vLLM with
+
+    .. code-block:: console
+
+        $ python setup.py develop
+
+    since it will give you incremental builds. The downside is that this method
+    is `deprecated by setuptools <https://github.com/pypa/setuptools/issues/917>`_.

setup.py

@@ -15,6 +15,11 @@
 
 ROOT_DIR = os.path.dirname(__file__)
 
+# If you are developing the C++ backend of vLLM, consider building vLLM with
+# `python setup.py develop` since it will give you incremental builds.
+# The downside is that this method is deprecated, see
+# https://github.com/pypa/setuptools/issues/917
+
 MAIN_CUDA_VERSION = "12.1"
 
 # Supported NVIDIA GPU architectures.