antgroup/glake

Implements virtual memory stitching and multi-path concurrent I/O to defragment GPU memory and accelerate CPU-GPU transfers without requiring code changes. Built as a pluggable layer beneath PyTorch that provides global memory pooling across GPUs, automatic deduplication for inference workloads, and specialized KV-cache optimization for LLM serving. Achieves 4× training throughput improvement and 3-12× I/O acceleration through memory defragmentation, tiering, and parallel transfer channels.

499 stars. No commits in the last 6 months.