AI Architecture & Systems

Training Memory Management

5.12.1Activation Checkpointing and Selective Recomputation#

5.12.2CPU/NVMe Offload and State Tiering#

5.12.3Activation Offload, Prefetch, and Asynchronous Transfer#

5.12.4Optimizer State Sharding and Low-Precision States#

5.12.5Memory Budget, Peak Memory, and Fragmentation#

5.12.6Training Graphs, Communication Buffers, and CUDA Graph Memory#