Crucible, a new distributed training platform, combines SparseLoco with subspace network compression to enable LLM training across bandwidth-constrained, geographically dispersed clusters. The paper addresses fault tolerance by using bucket storage for resilient all-gather operations and adapting SkipPipe techniques to reroute computations around failed pipeline stages, allowing training to continue despite hardware faults.
Nvidia expanded its open-source CUDA-Q platform with a new Logical orchestration layer enabling researchers to design and test fault-tolerant quantum computing applications. Fermilab achieved a 7x speedup in algorithm development using the tool, reducing timelines from five months to three weeks. Sandia National Laboratories developed QUOPS, a benchmark for measuring quantum hardware readiness for practical applications, now integrated into CUDA-Q.
Baldur is a Python framework that captures failed function calls when dependencies go down and automatically replays them upon recovery, using decorators to add circuit breakers, retry logic, and dead-letter queue functionality with built-in web console, metrics, and adapters for Django, FastAPI, Flask, and Celery.
A curated collection of seminal research papers in distributed systems, spanning foundational work on event ordering, Byzantine fault tolerance, consensus algorithms, and replication methods from 1978 to 2014. The list includes influential contributions from researchers like Leslie Lamport and covers critical problems that shaped modern distributed computing.