Cartridges·STILL Single-GPU Reproduction

An open-source implementation reproducing Cartridges and STILL on a single GPU has been released.

• cartridges reproduces corpus-specific compressed KV caches.
• STILL reproduces reusable neural KV-cache compaction.
• The STILL repository also includes benchmarks comparing full-context inference, truncation, and cartridges.
• The goal is not to summarize papers or blog posts, but to let people directly check and run things through readable implementations and benchmark code.

The linked original sources are the Cartridges paper (arXiv) and the STILL research post (Baseten), respectively.