HeadInfer: Memory-Efficient LLM Inference by Head-wise Offloading
This addresses the memory constraints for deploying LLMs on resource-limited hardware, though it is an incremental improvement over existing offloading techniques.
The paper tackles the memory bottleneck in large language model inference caused by key-value cache storage by proposing HEADINFER, a head-wise offloading strategy that reduces GPU memory usage for the KV cache from 128 GB to 1 GB on a 1-million-token sequence with Llama-3-8B, enabling 4-million-token inference on a single consumer GPU.
Transformer-based large language models (LLMs) demonstrate impressive performance in long context generation. Extending the context length has disproportionately shifted the memory footprint of LLMs during inference to the key-value cache (KV cache). In this paper, we propose HEADINFER, which offloads the KV cache to CPU RAM while avoiding the need to fully store the KV cache for any transformer layer on the GPU. HEADINFER employs a fine-grained, head-wise offloading strategy, maintaining only selective attention heads KV cache on the GPU while computing attention output dynamically. Through roofline analysis, we demonstrate that HEADINFER maintains computational efficiency while significantly reducing memory footprint. We evaluate HEADINFER on the Llama-3-8B model with a 1-million-token sequence, reducing the GPU memory footprint of the KV cache from 128 GB to 1 GB and the total GPU memory usage from 207 GB to 17 GB, achieving a 92% reduction compared to BF16 baseline inference. Notably, HEADINFER enables 4-million-token inference with an 8B model on a single consumer GPU with 24GB memory (e.g., NVIDIA RTX 4090) without approximation methods.