python3Packages.vllm: unvendor pynvml

- Replaces `pynvml` dependency with `nvidia-ml-py`
- Patches vLLM to use system `pynvml` from `nvidia-ml-py` instead of its vendored copy

Removes need to manually set `LD_LIBRARY_PATH=/run/opengl-driver/lib` when running `vllm` commands, as `nvidia-ml-py` includes patches to locate `libnvidia-ml.so.1` in Nix environments.

pkgs/development/python-modules/vllm/default.nix

@@ -34,7 +34,7 @@
   uvicorn,
   pydantic,
   aioprometheus,
-  pynvml,
+  nvidia-ml-py,
   openai,
   pyzmq,
   tiktoken,
@@ -314,6 +314,13 @@ buildPythonPackage rec {
   ];
 
   postPatch = ''
+    # Remove vendored pynvml entirely
+    rm vllm/third_party/pynvml.py
+    substituteInPlace tests/utils.py \
+      --replace-fail "from vllm.third_party.pynvml import" "from pynvml import"
+    substituteInPlace vllm/utils/__init__.py \
+      --replace-fail "import vllm.third_party.pynvml" "import pynvml"
+
     # pythonRelaxDeps does not cover build-system
     substituteInPlace pyproject.toml \
       --replace-fail "torch ==" "torch >=" \
@@ -442,7 +449,7 @@ buildPythonPackage rec {
   ]
   ++ lib.optionals cudaSupport [
     cupy
-    pynvml
+    nvidia-ml-py
     flashinfer
   ];