Dynamic Tensor Rematerialization

Checkpointing enables training deep learning models under restricted memory budgets by freeing intermediate activations from memory and recomputing them on demand. Previous checkpointing techniques statically plan these recomputations offline and assume static computation graphs. We demonstrate that a simple online algorithm can achieve comparable performance by introducing Dynamic Tensor Rematerialization (DTR), a greedy online algorithm for checkpointing that is extensible and general, is parameterized by eviction policy, and supports dynamic models. We prove that DTR can train an N-layer linear feedforward network on an Ω(N) memory budget with only O(N) tensor operations. DTR closely matches the performance of optimal static checkpointing in simulated experiments. We incorporate a DTR prototype into PyTorch just by interposing on tensor allocations and operator calls and collecting lightweight metadata on tensors.