Large language models (LLMs) based on transformer architecture have shown outstanding performance across numerous real-world tasks. However, the autoregressive nature of these models makes the inference process slow and costly. Speculative decoding has emerged as a promising solution, leveraging a smaller auxiliary model to draft future tokens, which are then validated simultaneously by the larger model, achieving a speed-up of 1-2x. Although speculative decoding matches the same distribution as multinomial sampling, multinomial sampling itself is prone to suboptimal outputs, where as beam sampling is widely recognized for producing higher-quality results by maintaining multiple candidate sequences at each step. This paper explores the novel integration of speculative decoding with beam sampling. However, there are four key challenges: (1) how to generate multiple sequences from the larger model's distribution given drafts sequences from the small model; (2) how to dynamically optimize the number of beams to balance efficiency and accuracy; (3) how to efficiently verify the multiple drafts in parallel; and (4) how to address the extra memory costs inherent in beam sampling. To address these challenges, we propose dynamic-width speculative beam decoding (DSBD). Specifically, we first introduce a novel draft and verification scheme that generates multiple sequences following the large model's distribution based on beam sampling trajectories from the small model. Then, we introduce an adaptive mechanism to dynamically tune the number of beams based on the context, optimizing efficiency and effectiveness. Besides, we extend tree-based parallel verification to handle multiple trees simultaneously, accelerating the verification process. Finally, we illustrate a simple modification to our algorithm to mitigate the memory overhead of beam sampling. Experimental results show that our approach achieves a 1.5-1.9x speed-up and1.8-2.5x lower energy consumption compared to beam sampling, with no loss in downstream performance. Moreover, it can produce significantly higher-quality outputs than speculative decoding, while maintaining similar time, memory, and energy costs. In summary, our method offers a more efficient and effective inference process for LLMs.

Large language models (LLMs) have achieved remarkable success across diverse tasks, yet their inference processes are hindered by substantial time and energy demands due to single-token generation at each decoding step. While previous methods such as speculative decoding mitigate these inefficiencies by producing multiple tokens per step, each token is still generated by its single-token distribution, thereby enhancing speed without improving effectiveness. In contrast, our work simultaneously enhances inference speed and improves the output effectiveness. We consider multi-token joint decoding (MTJD), which generates multiple tokens from their joint distribution at each iteration, theoretically reducing perplexity and enhancing task performance. However, MTJD suffers from the high cost of sampling from the joint distribution of multiple tokens. Inspired by speculative decoding, we introduce multi-token assisted decoding (MTAD), a novel framework designed to accelerate MTJD. MTAD leverages a smaller auxiliary model to approximate the joint distribution of a larger model, incorporating a verification mechanism that not only ensures the accuracy of this approximation, but also improves the decoding efficiency over conventional speculative decoding. Theoretically, we demonstrate that MTAD closely approximates exact MTJD with bounded error. Empirical evaluations using Llama-2 and OPT models ranging from 13B to 70B parameters across various tasks reveal that MTAD reduces perplexity by 21.2% and improves downstream performance compared to standard single-token sampling. Furthermore, MTAD achieves a 1.42× speed-up and consumes 1.54× less energy than conventional speculative decoding methods. These results highlight MTAD’s ability to make multi-token joint decoding both effective and efficient, promoting more sustainable and high-performance deployment of LLMs.

Zifan He, Yingqi Cao, Zongyue Qin, Neha Prakriya, Yizhou Sun, Jason Cong. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Despite the increasing adoption of FPGAs in compute clouds, there remains a significant gap in programming tools and abstractions which can leverage network-connected, cloud-scale, multi-die FPGAs to generate accelerators with high frequency and throughput. We propose TAPA-CS, a task-parallel dataflow programming framework which automatically partitions and compiles a large design across a cluster of FPGAs while achieving high frequency and throughput. TAPA-CS has three main contributions. First, it is an open-source framework which allows users to leverage virtually "unlimited" accelerator fabric, high-bandwidth memory (HBM), and on-chip memory. Second, given as input a large design, TAPA-CS automatically partitions the design to map to multiple FPGAs, while ensuring congestion control, resource balancing, and overlapping of communication and computation. Third, TAPA-CS couples coarse-grained floor-planning with interconnect pipelining at the inter- and intra-FPGA levels to ensure high frequency. FPGAs in our multi-FPGA testbed communicate through a high-speed 100Gbps Ethernet infrastructure. We have evaluated the performance of TAPA-CS on designs, including systolic-array based CNNs, graph processing workloads such as page rank, stencil applications, and KNN. On average, the 2-, 3-, and 4-FPGA designs are 2.1×, 3.2×, and 4.4× faster than the single FPGA baselines generated through Vitis HLS. TAPA-CS also achieves a frequency improvement between 11%-116% compared with Vitis HLS.

The identification and quantification of proteins through mass spectrometry (MS) are foundational to proteomics, offering insights into biological systems and disease states. However, current clustering tools struggle to process large-scale datasets. We propose SpectraFlux, a multiple FPGA-based architecture for accelerated mass spectrum clustering that outperforms existing CPU, GPU, and FPGA designs. It employs heterogeneous clustering kernels for adaptive bucket size management and optimizes memory usage by distinguishing between on-chip and high-bandwidth memory (HBM) storage solutions. SpectraFlux is built upon the TAPA-CS framework, which automatically compiles and partitions a large dataflow design across multiple chips with RDMA-based inter-FPGA communication. Our solution shows a 2.7X speed up on a quad-FPGA platform compared to a single FPGA. Additionally, we introduce a refined cost model for frame-based inter-FPGA communication to better accommodate the variable data rates inherent in proteomic data processing. This reduces the inter-FPGA data movement by up to 73%. Finally, SpectraFlux achieves speedups of up to 11X and 17X over SOTA FPGA and GPU accelerators, respectively.

Large-scale machine learning (ML) models rely on extremely large datasets to learn their exponentially growing number of parameters. While these models achieve unprecedented success, the increase in training time and hardware resources required is unsustainable. Further, we find that as dataset sizes increase, data movement becomes a significant component of overall training time. We propose NeSSA, a novel SmartSSD+GPU training architecture to intelligently select important subsets of large datasets near-storage, such that training on the subset mimics training on the full dataset with a very small loss in accuracy. To the best of our knowledge, this is the first work to propose such a near-storage data selection model for efficient ML training. We have evaluated our method for the CIFAR-10, SVHN, CINIC-10, CIFAR-100, TinyImageNet, and ImageNet-100 datasets. We also test across ResNet-20, ResNet-18, and ResNet-50 models.