# wheels.baonq.me

> Read-only mirror of Python ML wheels, NVIDIA toolkits, and Linux boot images for aarch64 and x86_64 hosts.

Every directory has a Markdown twin: append `.md` to any directory path, or use
`/index.md` for the root. The twin lists names, types, sizes, and modification
times as a Markdown table. Links in it are absolute paths, and a subdirectory
links to its own twin, so the whole tree can be walked with plain GETs and no
HTML parsing. Files themselves are served unchanged at their normal paths and
support byte ranges, so large downloads resume.

Directory listings also answer `Accept: application/json` if you would rather
parse JSON than Markdown.

The wheel trees are also a PEP 503 package index at
https://wheels.baonq.me/simple/, so pip can install from them directly with
`pip install --index-url https://wheels.baonq.me/simple/ torch` rather than by
fetching a file by URL. It answers the PEP 691 JSON API too, for clients that
send `Accept: application/vnd.pypi.simple.v1+json`.

## Index
- [Archive root](https://wheels.baonq.me/index.md): Top-level areas — Python wheels, CUDA and cuDNN installers, cloud images, installer ISOs

## Python wheels
- [Python wheels](https://wheels.baonq.me/wheels.md): Entry point to the wheel trees, one directory per CPU architecture and microarchitecture
- [x86_64 wheels](https://wheels.baonq.me/wheels/x86_64.md): torch, torchvision and vLLM for x86_64, CUDA 12.8 and 12.9, including a CPU-only vLLM
- [x86_64 torch, by Ubuntu base](https://wheels.baonq.me/wheels/x86_64/torch.md): torch 2.6.0 for x86_64 only, split by CUDA version and Ubuntu base image
- [x86_64 torchvision](https://wheels.baonq.me/wheels/x86_64/torchvision.md): One torchvision 0.21.0 wheel for x86_64, built against CUDA 12.8
- [Generic aarch64 wheels](https://wheels.baonq.me/wheels/aarch64.md): torch 2.9 to 2.10 and vLLM 0.13 to 0.16 for any aarch64, CUDA 12.9 and 13.0
- [ARMv8 wheels](https://wheels.baonq.me/wheels/aarch64-armv8.md): ARMv8 serving stack — torch 2.5 to 2.10, vLLM, sglang, flashinfer, triton, lmcache
- [ARMv9 wheels](https://wheels.baonq.me/wheels/aarch64-armv9.md): Largest tree — ARMv9 torch, torchaudio, torchvision, triton and vLLM, CUDA 12.8 to 13.2
- [vLLM nightly aarch64 wheels](https://wheels.baonq.me/wheels/aarch64-armv9/vllm_latest.md): Daily unofficial vLLM main builds for aarch64 CUDA, each with a JSON sidecar and a README

## NVIDIA toolkits
- [CUDA toolkits](https://wheels.baonq.me/cuda.md): CUDA toolkit runfile installers, one directory per target architecture
- [CUDA for x86_64](https://wheels.baonq.me/cuda/x86_64.md): CUDA 11.8 through 13.2 .run installers for x86_64, with their bundled driver versions
- [CUDA for arm64 SBSA](https://wheels.baonq.me/cuda/arm64-sbsa.md): CUDA 11.8 through 13.2 .run installers for arm64 SBSA servers such as Grace
- [cuDNN](https://wheels.baonq.me/cudnn.md): cuDNN 9.14 local-repo .deb packages and install scripts for Ubuntu 22.04 and 24.04
- [NVIDIA HPC SDK](https://wheels.baonq.me/hpc.md): NVIDIA HPC SDK 2025 tarball for aarch64, built against CUDA 13.0

## OS images
- [Cloud images](https://wheels.baonq.me/cloudimg.md): Cloud images grouped by distribution, all arm64, qcow2 and raw img formats
- [Ubuntu cloud images](https://wheels.baonq.me/cloudimg/ubuntu.md): Ubuntu arm64 server cloud images, trusty through noble, raw img format
- [Debian cloud images](https://wheels.baonq.me/cloudimg/debian.md): Debian arm64 generic cloud images, versions 9 through 12, qcow2 format
- [Rocky Linux cloud images](https://wheels.baonq.me/cloudimg/rocky.md): Rocky Linux aarch64 GenericCloud base images for releases 8, 9 and 10
- [CentOS cloud images](https://wheels.baonq.me/cloudimg/centos.md): CentOS 7 plus CentOS Stream 8, 9 and 10 aarch64 GenericCloud images
- [Installer ISOs](https://wheels.baonq.me/iso.md): Ubuntu 24.04 installer ISOs, amd64 desktop and arm64 live server

## Optional
- [Mirror scripts](https://wheels.baonq.me/mirror.md): Scripts that populate the wheel trees from upstream PyTorch download indexes