thinking-machines-lab/batch_invariant_ops (GitHub repository)
Thinking Machines Lab. 2025. GitHub.
See what cites it| Link | https://github.com/thinking-machines-lab/batch_invariant_ops |
|---|---|
| Version | Main branch as viewed on 2026-09-24 (7 commits). MIT licence. |
| Accessed | 2026-09-24 |
| Note | Companion library to S-1009. Replaces torch.mm, torch.addmm, torch.log_softmax and torch.mean with batch-invariant kernels through torch.Library, and includes a deterministic vLLM example: 18 unique samples out of 1000 completions without the upstream vLLM change, one unique sample with it. |