LATTICE

A persistent memory library for efficient ML model versioning that enables low-latency direct persistence checkpointing on non-volatile memory devices, published at SYSTOR 2025.

ai-systems active
cpythoncudapytorchdarknetpmdk
Source Code → Paper →

LATTICE

LATTICE is a persistent memory library that enables efficient in-memory DNN model versioning. By leveraging non-volatile memory expansion devices for direct persistence, it bypasses the traditional file system bottleneck in ML checkpointing workflows.

Key Results

  • Low-latency checkpoint writes through direct persistence on non-volatile memory
  • Efficient model versioning that maintains multiple checkpoint snapshots without redundant storage
  • Integrated with the Darknet deep learning framework for real-world validation
  • Published at SYSTOR 2025 (ACM International Systems and Storage Conference)