NVIDIA/kvpress v0.4.0

v0.4.0

Repository: NVIDIA/kvpress

Tag: v0.4.0

Published: 2025-12-05T08:54:36Z

Prerelease: no

Release notes:

🚀 Release v0.4.0

✨ New Features

• CURPress - Value-Guided KV Compression for LLMs via Approximated CUR Decomposition (#150)
• CompactorPress - Compactor: Calibrated Query-Agnostic KV Cache Compression with Approximate Leverage Scores (#143)
• Decoding Press Functionality - Support for KV cache compression during the decoding phase (#139)
• AIME25 & Math500 Benchmarks - New evaluation datasets for mathematical reasoning tasks (#142)
• `post_init_from_model` Hook - Add model-specific initialization support in BasePress (#163)

📈 Improvements

• Moved tests to GPU for faster CI execution (#132)
• Improved needle-in-haystack test coverage (#133)
• Updated README and documentation for clarity (#162)
• Enhanced docstrings throughout the codebase (#159)
• Updated decoding notebook with latest examples (#156)
• Code cleanup: moved utilities, cleaned imports (#160)

🐛 Bug Fixes

• Fixed LongBench-v2 benchmark evaluation (#161)
• Fixed kvzip press access to past_key_values
• Fixed ComposedPress behavior (#148)
• Fixed import issues (#144)

📦 Installation

pip install kvpress==0.4.0

📚 Full Changelog

https://github.com/NVIDIA/kvpress/compare/v0.3.0...v0.4.0