In long-context large language model (LLM) inference, the prefill stage dominates computation due to self-attention over the complete input context. Sparse attention significantly reduces self-attention computation by limiting each token's interactions to a subset of tokens. The attention sparsity pattern varies across input prompts, and within a prompt, each attention head can follow a distinct pattern. This makes attention sparsity dynamic. The requirement of generating the sparsity pattern, combined with limited data reuse in attention, shifts the prefill compute to being memory-bound. This, in addition to the huge energy requirements for long-context inference on GPU, motivates FPGAs as good candidates for accelerating dynamic long-context inference. To tackle these challenges, we propose FAST-Prefill, the first FPGA accelerator for long-context prefill-stage inference with dynamic sparse attention. To efficiently generate sparse indices, we propose a \textit{fused pipeline unit with a memory-aware execution order} to reduce large tensors and irregular memory accesses. To reduce off-chip memory traffic for accessing the KV cache, we utilize the memory hierarchy to design a \textit{liveness-driven, dual-tier cache}. For high-throughput matrix multiplication, we design a \textit{hybrid Matrix Processing Unit (MPU)} with DSPs and bit-plane decomposition using LUTs. We implement FAST-Prefill on Alveo U280 and evaluate it on the Llama and Qwen models (batch size = 1) for context lengths ranging from 4K to 128K tokens. We demonstrate an average speedup of up to 2.5$\times$ in TTFT and 4.5$\times$ improvement in energy efficiency over GPU implementation on Nvidia A5000 GPU.

翻译：暂无翻译