SkyRL prebuilt CUDA extension wheels — torch 2.13 / CUDA 13.0

Stopgap wheels so SkyRL can move to torch 2.13.

Why

SkyRL pins several CUDA extensions from Astral's GPU wheel index, which tags each build with the exact torch it was compiled against (+cu.13.0.torch.2.11). As of 2026-09-03 that index has no torch 2.13 builds — flash-attn, causal-conv1d and transformer-engine-torch stop at torch.2.12, and mamba-ssm at torch.2.11. So uv lock cannot resolve a torch 2.13 environment.

torch 2.13 is wanted because vLLM 0.28's vendored DeepGEMM extension is not stable-ABI and is linked against torch 2.13's c10; under torch 2.11 it fails with undefined symbol: c10::ValueError(c10::SourceLocation, std::string), disabling the DeepGEMM fused-MoE and sparse-attention-indexer paths.

These wheels are temporary. Delete them and go back to the Astral index as soon as it publishes torch.2.13 builds.

Build environment

torch	2.13.0+cu130
CUDA	13.0.88 (cuda-toolkit-13-0, 13.0.3-1)
Python	3.12 (cp312 only)
TORCH_CUDA_ARCH_LIST	9.0;10.0 — H100 and B200 only
OS	Ubuntu 24.04, x86_64, glibc 2.39

Note on architecture coverage

Astral's wheels cover many more architectures. These cover sm_90 (H100) and sm_100 (B200) only, to keep build times sane. On any other GPU they will fail at kernel launch. That is the main reason not to treat these as a general replacement.

flash-attn is built at exactly 2.8.3, not 2.8.3.post1: Transformer Engine gates flash-attn on max_version = 2.8.3, and the .post1 compares greater.

Verification

Each wheel was installed into a clean torch 2.13.0+cu130 venv and exercised on a B200 — import of the compiled extension plus a forward pass returning finite values.

Reproducing

See build_all.sh in this repo.