qwen35-cu13-on12

Frozen Python virtual environment used to serve the Qwen3.5 / Qwen3.6 family with vLLM in the observation-masking project.

The repo name qwen35-cu13-on12 reflects the setup: CUDA 13 runtime libraries (cu13) running on a CUDA 12.8 PyTorch build / driver host (on12), via the bundled cuda13-compat/ shim. This combination is what let the nightly vLLM build run Qwen3.5/3.6 on hosts that did not yet ship a full CUDA 13 toolkit.

⚠️ A raw venv is not relocatable. pyvenv.cfg and many bin/ scripts hardcode the original absolute path, and the tree contains compiled binaries and symlinks. Do not expect to download and source activate this directly. Treat it as an exact archive of installed versions. To reproduce the environment, use requirements.txt (see below).

Key versions

Component	Version
Python	3.12.13
vLLM	0.21.0
PyTorch	2.11.0+cu128
torchvision / torchaudio	0.26.0+cu128 / 2.11.0+cu128
Triton	3.6.0
FlashInfer	0.6.8.post1
Transformers	5.12.1
nvidia-cuda-runtime	13.3.29 (cu13) + 12.8.90 (cu12)
nvidia-cuda-nvrtc	13.3.33 (cu13) + 12.8.93 (cu12)
nvidia-cudnn-cu12	9.19.0.56
nvidia-nccl-cu12	2.28.9

Full pinned list: requirements.txt (187 packages).

CUDA 13 compatibility shim (cuda13-compat/)

The cuda13-compat/ directory carries the CUDA 13 runtime libraries that the nightly vLLM build links against, alongside the host's NVIDIA driver libraries (driver 580.167.08). It is added to LD_LIBRARY_PATH at launch so a cu13-linked vLLM can run on a cu12 PyTorch / driver host.

libcuda.so -> libcuda.so.1 -> libcuda.so.580.167.08
libcudadebugger.so.1 -> libcudadebugger.so.580.167.08
libcudart.so.13
libnvidia-gpucomp.so.580.167.08
libnvidia-nvvm.so -> libnvidia-nvvm.so.4 -> libnvidia-nvvm.so.580.167.08
libnvidia-nvvm70.so.4
libnvidia-pkcs11-openssl3.so.580.167.08
libnvidia-ptxjitcompiler.so.1 -> libnvidia-ptxjitcompiler.so.580.167.08
libnvrtc.so.13

Full listing with sizes: cuda13-compat.listing.txt.

How to reproduce (recommended)

Rather than reusing this venv, rebuild it with uv as in the project README:

uv venv qwen35 --python 3.12
uv pip install vllm \
  --torch-backend=auto \
  --extra-index-url https://wheels.vllm.ai/nightly \
  --python qwen35/bin/python

# then pin exact versions
uv pip install -r requirements.txt --python qwen35/bin/python

Citation

@article{zhang2026masking,
  title={Masking Stale Observations Helps Search Agents -- Until It Doesn't: A Regime Map and Its Mechanism},
  author={Zhang, Haoxiang and Xu, Qixin and Li, Zhuofeng and Zhang, Lei and Jiang, Pengcheng and Zhang, Yu and McAuley, Julian},
  journal={arXiv preprint arXiv:2606.00408},
  year={2026}
}