ba2d6ac8d8
The CUDA torch pin stays < 2.11 because flash-attn, causal-conv1d, and mamba-ssm do not publish torch 2.11 wheels yet; bumping now would drop those prebuilt accelerators and force slow source builds. This adds a version-compat canary that queries the GitHub releases of all three and fails only once all of them ship torch 2.11 Linux wheels, which is when it is safe to raise the pin. Green while any is pending, skips on a network error, and wired into version-compat-ci (PR paths + daily cron). |
||
|---|---|---|
| .. | ||
| ISSUE_TEMPLATE | ||
| scripts | ||
| workflows | ||
| CODEOWNERS | ||
| dependabot.yml | ||
| FUNDING.yml | ||