SlimInfer: Accelerating Long-Context LLM Inference via Dynamic Token Pruning
This addresses efficiency bottlenecks for users deploying long-context LLMs, though it is incremental as it builds on existing attention optimization methods.
The paper tackles the high computational demands of long-context LLM inference by proposing SlimInfer, a framework that prunes less critical prompt tokens during the forward pass, achieving up to 2.53x speedup in time-to-first-token and 1.88x reduction in end-to-end latency for LLaMA3.1-8B-Instruct without performance loss on LongBench.
Long-context inference for Large Language Models (LLMs) is heavily limited by high computational demands. While several existing methods optimize attention computation, they still process the full set of hidden states at each layer, limiting overall efficiency. In this work, we propose SlimInfer, an innovative framework that aims to accelerate inference by directly pruning less critical prompt tokens during the forward pass. Our key insight is an information diffusion phenomenon: As information from critical tokens propagates through layers, it becomes distributed across the entire sequence. This diffusion process suggests that LLMs can maintain their semantic integrity when excessive tokens, even including these critical ones, are pruned in hidden states. Motivated by this, SlimInfer introduces a dynamic fine-grained pruning mechanism that accurately removes redundant tokens of hidden state at intermediate layers. This layer-wise pruning naturally enables an asynchronous KV cache manager that prefetches required token blocks without complex predictors, reducing both memory usage and I/O costs. Extensive experiments show that SlimInfer can achieve up to $\mathbf{2.53\times}$ time-to-first-token (TTFT) speedup and $\mathbf{1.88\times}$ end-to-end latency reduction for LLaMA3.1-8B-Instruct on a single RTX 4090, without sacrificing performance on LongBench. Our code will be released upon acceptance.