v0.19.0-devnen.2 — Windows PP=2 fix

Windows multi-process executor fix (Ampere + Blackwell)

This is the unified release containing patched wheels for both Ampere/Ada (sm_86/89) and Blackwell (sm_120) Windows targets. Both wheels add the Windows-only fixups required for any multi-process executor path (PP>=2, multi-engine) to boot:

1. vllm/utils/network_utils.py — get_open_zmq_ipc_path() returns tcp://127.0.0.1:<port> on Windows. pyzmq has no ipc:// transport on Windows; without this, vLLM v1 EngineCore <-> APIServer <-> Worker channels crash at boot with ZMQError: Protocol not supported (addr='ipc://...'). New in this release for both variants.
2. vllm/v1/executor/multiproc_executor.py — widens isinstance(pipe, Connection) to isinstance(pipe, _ConnectionBase) so PipeConnection (Windows) is accepted alongside Connection (POSIX). New in this release for the Ampere wheel; the Blackwell wheel inherits this from the upstream 0.20.0 source.

Together these unblock PP=2 (and any other multi-process executor path) on Windows. Verified on a 2× RTX 3090 box running qwen3.6-windows-server pp2_160k.

Files

Variant	Wheel	Target
Ampere/Ada	vllm-0.19.0+devnen.2-cp312-cp312-win_amd64.whl	sm_86/89, CUDA 12.6, Python 3.12
Blackwell	vllm-0.20.0+cu132.devnen.2-cp312-cp312-win_amd64.whl	sm_120, CUDA 13.2, Python 3.12
	SHA256SUMS.txt	combined

Inherited from +devnen.1

• Ampere: CPU-relay for Gloo collectives (no NCCL on Windows), Qwen3 reasoning parser PR #35687 mirror, hardwired wildcard model-name acceptance.
• Blackwell: Qwen3 reasoning parser PR #35687 mirror, hardwired wildcard model-name acceptance. CPU-relay was dropped in 0.20.0 because it ships with experimental NCCL-on-Windows support.

See windows_patches/README.md for the full patch list.