Guanghui Zhang

Haofei Yin, Mengbai Xiao, Tinghong Li et al.

The demand for large language model inference is rapidly increasing. Pipeline parallelism offers a cost-effective deployment strategy for distributed inference but suffers from high service latency. While incorporating speculative decoding to pipeline parallelism improves performance, it still faces challenges of low hardware utilization and narrow speculative window. Inspired by branch prediction in instruction pipelining, we introduce SpecPipe, which fills the pipeline with speculative tokens of a request step-by-step. By maximizing the hardware utilization, SpecPipe decodes one token per pipeline step ideally. Specifically, SpecPipe comprises a dynamic speculative token tree and a pipelined inference framework. The tree dynamically accepts tokens from a speculative token source and outputs the tokens to the inference pipeline. Since the speculative window relaxed in our framework, a high-accuracy draft model is integrated without fine-tuning. The pipeline inference framework follows node-wise computation, pruning propagation, and inter-node communication stages. We implement SpecPipe and a variant SpecPipe-DB with dynamic batching for single- and multi-request inference, respectively. On an 8-stage pipeline, SpecPipe improves time between tokens on diverse single-request workloads by $4.19\times$-$5.53\times$ over standard pipeline parallelism and by $2.08\times$-$2.38\times$ over prior tree-based speculative decoding methods. For multi-request workloads, SpecPipe-DB achieves $1.64\times$-$2.08\times$ higher throughput and $1.61\times$-$2.06\times$ lower time between tokens than vLLM.

Haolin Jin, Mengbai Xiao, Yuan Yuan et al.

The Transformer architecture has revolutionized deep learning, delivering the state-of-the-art performance in areas such as natural language processing, computer vision, and time series prediction. However, its core component, self-attention, has the quadratic time complexity relative to input sequence length, which hinders the scalability of Transformers. The exsiting approaches on optimizing self-attention either discard full-contextual information or lack of flexibility. In this work, we design DistrAttention, an effcient and flexible self-attention mechanism with the full context. DistrAttention achieves this by grouping data on the embedding dimensionality, usually referred to as $d$. We realize DistrAttention with a lightweight sampling and fusion method that exploits locality-sensitive hashing to group similar data. A block-wise grouping framework is further designed to limit the errors introduced by locality sensitive hashing. By optimizing the selection of block sizes, DistrAttention could be easily integrated with FlashAttention-2, gaining high-performance on modern GPUs. We evaluate DistrAttention with extensive experiments. The results show that our method is 37% faster than FlashAttention-2 on calculating self-attention. In ViT inference, DistrAttention is the fastest and the most accurate among approximate self-attention mechanisms. In Llama3-1B, DistrAttention still achieves the lowest inference time with only 1% accuray loss.