Skip to content

vLLM patched native-Windows wheels (0.19.0 Ampere + 0.20.0 Blackwell)

Choose a tag to compare

View all tags
@devnen devnen released this 29 Apr 15:10
· 9 commits to main since this release
v0.19.0-devnen.1
bea51a8

vLLM patched native-Windows wheels

This release ships two wheels to cover both NVIDIA generations on
native Windows:

Wheel vLLM torch / CUDA GPUs Notes
vllm-0.19.0+devnen.1-cp312-cp312-win_amd64.whl 0.19.0 torch 2.7 / cu126 RTX 30/40 series (Ampere/Ada, sm_86/sm_89) Repackages SystemPanic/vllm-windows 0.19.0 with the three Windows patches listed below.
vllm-0.20.0+cu132.devnen.1-cp312-cp312-win_amd64.whl 0.20.0 torch 2.9 / cu132 RTX 50 series (Blackwell, sm_120) Driver 596+ required. Built against CUDA 13. End-users get the runtime via the qwen3.6-windows-server launcher's auto-built cuda13_shim.

Pick the one matching your GPU. Python 3.12, win_amd64 only.

What's patched (vs upstream vLLM)

  1. CPU-relay for Gloo collectives. Windows has no NCCL. PP/TP collectives hang or 0xC0000005 on CUDA tensors. Patches parallel_state.py, cuda_communicator.py, base_device_communicator.py, and gpu_worker.py to stage through pinned CPU buffers when os.name == "nt".
  2. Qwen3 reasoning parser fix. Mirror of upstream PR #35687.
  3. Hardwired wildcard model name. OpenAIModelRegistry.is_base_model always returns True; clients no longer need to match --served-model-name.

Full diff: CHANGES_VS_SYSTEMPANIC.md.

Install (developer / direct)

python -m venv venv
.\venv\Scripts\Activate.ps1
pip install <url-to-wheel-matching-your-GPU>

SHA256

8f537f97a9fb00c0504ca644671c13a3df33ccf663ba5c40612f388c79dc4471  vllm-0.19.0+devnen.1-cp312-cp312-win_amd64.whl
1ef872448f1c850e48f417cc89cb6394b9e7f01edbc8f186574a7fb6c1b5d4e8  vllm-0.20.0+cu132.devnen.1-cp312-cp312-win_amd64.whl

Looking for a one-click launcher?

Use devnen/qwen3.6-windows-server
— portable zip, embedded Python, the matching wheel bundled in (Ampere or Blackwell variant auto-selected).

Compatibility

  • Windows 10 / 11 x64
  • Python 3.12 (cp312)
  • NVIDIA driver: 553+ for Ampere/Ada, 596+ for Blackwell
Assets 5
Loading