---
license: other
license_name: mixed-third-party-runtime-licenses
pretty_name: Portable vLLM runtime
tags:
- vllm
- inference
- cuda
- portable-runtime
---
# Portable vLLM 0.32.0 / CUDA 13.0

Release `vllm-0.32.0-cu130-20261028-r1`. Payload revision: `2f96debc23c2ae0ca333314587bf321b28c25f37`.
Standalone Python 3.12.12, vLLM 0.32.0, Torch 2.13.0, an exact `uv.lock`,
private CUDA 13.0 toolkit, support libraries and node helpers.
Linux x86_64, glibc >= 2.35. Fixed installation root:
`/opt/vllm-runtime-0.32.0-cu130` (root permission required).

Archive size: **5,821,293,851 bytes**.
Unpacked regular-file bytes: **11,041,486,896**.
Archive SHA256: `b22e2ef9b1d1f1f2f233b9360bc82e8cb53b69c3d4811e3043167f0037db532b`.
Reserve space for the downloaded archive, extraction, JIT caches, models and outputs.

## Download

Download [download_release.py](https://huggingface.co/datasets/codemaivanngu/vllm-portable-runtime/resolve/2f96debc23c2ae0ca333314587bf321b28c25f37/releases/vllm-0.32.0-cu130-20261028-r1/download_release.py).
With an existing Python 3.10+ on the company Linux instance:

```bash
python3 download_release.py --revision 2f96debc23c2ae0ca333314587bf321b28c25f37 --archive-sha256 b22e2ef9b1d1f1f2f233b9360bc82e8cb53b69c3d4811e3043167f0037db532b --output /your/node-local/vllm-payload --workers 8
```

Use your existing approved proxy environment. This public release requires no
HF token. The downloader verifies every manifest file; it neither installs nor
uses GPUs. From the payload directory, verify `sha256sum -c SHA256SUMS`, then
run `python3 install.py`. The installer refuses existing paths and runs CPU/import
checks with GPUs hidden. It changes no driver or system CUDA installation.

## Run

The foreground CLI wrapper is
`/opt/vllm-runtime-0.32.0-cu130/node/vllm.sh`; pass normal vLLM arguments such as
`serve /path/to/your/model --host 127.0.0.1 --port 8000`.
`node/python-vllm.sh` runs Python scripts in the same independent runtime.
Both wrappers default to GPUs hidden. Explicitly set `CUDA_VISIBLE_DEVICES`
to your assigned GPU UUID(s), following your scheduler's assignment.
Stopping the server and its owned workers releases their VRAM; shell
deactivation alone does not stop processes.

Production models/data/output and JIT caches remain external. HF defaults to
offline mode; stage weights separately. Override `VLLM_PORTABLE_CACHE` for the
cache directory and `VLLM_PORTABLE_HF_HOME` for the model cache.
The node still needs a compatible NVIDIA driver, host glibc and a C/C++
compiler with development headers (GCC 11 was tested). nvcc/CUDA headers/ptxas
come from the private bundled toolkit. No NVIDIA kernel driver or host glibc
is bundled. Third-party licenses and CUDA notices remain applicable; this
bundle does not relicense dependencies.

## Evidence and limits

Publication requires CPU checks, native B200 CUDA kernels, normal compiled
synthetic dense BF16 generation and loopback HTTP serving, plus archive restore
and repeated GPU inference/HTTP checks on a separate Ubuntu 22.04 image with
no system CUDA toolkit and Internet blocked. See `qualification.json`, both restore receipts,
`build-facts.json`, `manifest.json` and `uv.lock` under `releases/vllm-0.32.0-cu130-20261028-r1/`.

Fixtures are tiny random models, not production weights or benchmark results.
The company node, real model families, multi-GPU/NCCL, quantization, multimodal
and all optional backends have not been qualified by these tests.
The release contains no corporate data, credentials or run caches.
A full anonymous download of the archive and every manifest file passed SHA256 verification. The receipt is published under `validation/vllm-0.32.0-cu130-20261028-r1/public-download-receipt.json`.