Reproducible investigations, debugging notes, and the limits of their evidence.
- A 16-GPU KV cache failure, reduced to 4 MiB—and why a larger buffer was not enough — a 4 MiB native-allocator reproducer, the initial candidate, and what review changed.
Each note distinguishes observed results, proposed explanations, and unverified behavior. Results are tied to specific software revisions; an open PR is not a completed fix.
Code is available under Apache-2.0. Upstream-derived files retain their original copyright and license headers; see NOTICE.