November 2023 Summaries

Running inference for a 70B parameter large language model on a single 4GB GPU is made feasible by several advanced memory optimization techniques that avoid model performance sacrifices. Key strategies include layer-wise inference, where each transformer layer is individually loaded, executed, and then memory-freed, significantly reducing the memory footprint to about 1.6GB per layer; flash attention, which optimizes memory access by reducing memory complexity from O(n²) to O(n) for faster computations; and model file sharding, which minimizes disk reading by preprocessing model files to align with layer sizes rather than original large shards. Additionally, the use of a meta device feature allows dynamic model part transfers between devices during execution, maintaining minimal memory usage. The open-source library AirLLM, available on GitHub, implements these techniques and allows inference on lower-end GPUs, though not ideal for interactive applications, and the possibility of achieving similar memory efficiency for training through methods like gradient checkpointing is suggested.