-
fix: support RTX 50-series (Blackwell sm_120) CUDA via cu128 wheels (#217) (#218)
frostbyte_neo 发布于
2026-06-23 21:28:56 +00:00 | 217 次提交 在此版本后已推送到 main- fix: support RTX 50-series (Blackwell sm_120) CUDA via cu128 wheels
The stock torch 2.6 cu12x wheels have no sm_120 kernels, so Blackwell
GPUs (e.g. RTX 5060 Ti) pass torch.cuda.is_available() but crash mid-
extraction with "no kernel image is available for execution on the
device" (#217).Targeted carve-out so existing users are untouched:
- detect_nvidia_gpu now also reads the GPU compute capability
- wheel_tag routes sm_100/sm_120 (cap major >= 10) to cu128, everything
else keeps the existing cu124/cu121/cu118 heuristic - install_cuda_torch installs torch 2.7.1+cu128 for the cu128 tag and
stays on 2.6.0 for all other tags - verify_cuda_torch now forces a real kernel launch instead of trusting
is_available(), so an incompatible wheel falls back to CPU cleanly
instead of crashing during a job
Closes #217
- fix(deps): bump msgpack 1.1.2 -> 1.2.1 (GHSA-6v7p-g79w-8964)
Trivy flags msgpack 1.1.2 with a HIGH advisory (out-of-bounds read /
crash on Unpacker reuse). It's a transitive dep via the torch/demucs
stack; bump to the fixed 1.2.1 to clear the trivy fs scan.
Co-authored-by: Thales <>
下载附件