Zongyue Qin

Hanyu Li, Yi-Ping Hsu, Aditya Mantha et al. · stanford

Modern recommendation systems predominantly train retrieval and ranking as separate models despite both increasingly relying on large transformers encoding the same user behavior data, duplicating parameters, compute, and serving cost. Prior work unifies the model architecture but not the full pipeline: input formats, training procedures, and serving stacks remain fragmented across stages. We present UniPinRec, which achieves full-stack unification of retrieval and ranking at Pinterest: one input format, one model, one training stage, deployed within existing serving infrastructure. A shared transformer encodes the user action sequence into candidate-independent representations that branch into retrieval (ANN dot-product) and ranking (cross-attention) via task-specific heads. Three ideas make this work: (1) Masked Action Modeling (MAM) eliminates interleaving, enabling weight sharing without doubling context length; (2) Blended training examples pair action sequences with feedview impression slates to satisfy both objectives jointly; (3) Cross-stage KV cache sharing reuses user-history computation from retrieval for ranking, reducing total FLOPs versus serving two independent models. Deployed in the Pinterest core surfaces, UniPinRec delivers approximately +1% online engagement lift while cutting end-to-end serving latency by 11.1% and lifting QPS by 63.6%. To our knowledge, this is the first full-stack unification of retrieval and ranking, covering inputs, model, training and serving, deployed in a production recommendation system.

Zongyue Qin, Ziniu Hu, Zifan He et al.

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 1.42x speed-up and consumes 1.54x 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.

Zongyue Qin, Raghavv Goel, Mukul Gagrani et al.

Speculative decoding has emerged as a powerful approach to accelerate large language model (LLM) inference by employing lightweight draft models to propose candidate tokens that are subsequently verified by the target model. The effectiveness of this paradigm critically depends on the quality of the draft model. While recent advances such as the EAGLE series achieve state-of-the-art speedup, existing draft models remain limited by error accumulation: they condition only on the current prefix, causing their predictions to drift from the target model over steps. In this work, we propose \textbf{ConFu} (Contemplate the Future), a novel speculative decoding framework that enables draft models to anticipate the future direction of generation. ConFu introduces (i) contemplate tokens and soft prompts that allow the draft model to leverage future-oriented signals from the target model at negligible cost, (ii) a dynamic contemplate token mechanism with MoE to enable context-aware future prediction, and (iii) a training framework with anchor token sampling and future prediction replication that learns robust future prediction. ConFu improves token acceptance rates and generation speed over EAGLE-3 by 8--11\% on Llama-3 3B/8B and by approximately 20\% on Qwen-3 4B across downstream tasks. We believe our work is the first to bridge speculative decoding with continuous reasoning tokens, offering a new direction for accelerating LLM inference.

Zongyue Qin, Zifan He, Neha Prakriya et al.

Large language models (LLMs) 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, whereas 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...

Zijian Ding, Atefeh Sohrabizadeh, Weikai Li et al.

There have been several recent works proposed to utilize model-based optimization methods to improve the productivity of using high-level synthesis (HLS) to design domain-specific architectures. They would replace the time-consuming performance estimation or simulation of design with a proxy model, and automatically insert pragmas to guide hardware optimizations. In this work, we address the challenges associated with high-level synthesis (HLS) design space exploration (DSE) through the evolving landscape of HLS tools. As these tools develop, the quality of results (QoR) from synthesis can vary significantly, complicating the maintenance of optimal design strategies across different toolchains. We introduce Active-CEM, a task transfer learning scheme that leverages a model-based explorer designed to adapt efficiently to changes in toolchains. This approach optimizes sample efficiency by identifying high-quality design configurations under a new toolchain without requiring extensive re-evaluation. We further refine our methodology by incorporating toolchain-invariant modeling. This allows us to predict QoR changes more accurately despite shifts in the black-box implementation of the toolchains. Experiment results on the HLSyn benchmark transitioning to new toolchain show an average performance improvement of 1.58$\times$ compared to AutoDSE and a 1.2$\times$ improvement over HARP, while also increasing the sample efficiency by 5.26$\times$, and reducing the runtime by 2.7$\times$.

Zongyue Qin, Chen Luo, Zhengyang Wang et al.

Large language models (LLMs) excel in many natural language processing (NLP) tasks. However, since LLMs can only incorporate new knowledge through training or supervised fine-tuning processes, they are unsuitable for applications that demand precise, up-to-date, and private information not available in the training corpora. This precise, up-to-date, and private information is typically stored in relational databases. Thus, a promising solution is to augment LLMs with the inclusion of relational databases as external memory. This can ensure the timeliness, correctness, and consistency of data, and assist LLMs in performing complex arithmetic operations beyond their inherent capabilities. However, bridging the gap between LLMs and relational databases is challenging. It requires the awareness of databases and data values stored in databases to select correct databases and issue correct SQL queries. Besides, it is necessary for the external memory to be independent of the LLM to meet the needs of real-world applications. We introduce a novel LLM-agnostic memory architecture comprising a database selection memory, a data value memory, and relational databases. And we design an elegant pipeline to retrieve information from it. Besides, we carefully design the prompts to instruct the LLM to maximize the framework's potential. To evaluate our method, we compose a new dataset with various types of questions. Experimental results show that our framework enables LLMs to effectively answer database-related questions, which is beyond their direct ability.

Weikai Li, Ding Wang, Zijian Ding et al.

High-level synthesis (HLS) is a widely used tool in designing Field Programmable Gate Array (FPGA). HLS enables FPGA design with software programming languages by compiling the source code into an FPGA circuit. The source code includes a program (called "kernel") and several pragmas that instruct hardware synthesis, such as parallelization, pipeline, etc. While it is relatively easy for software developers to design the program, it heavily relies on hardware knowledge to design the pragmas, posing a big challenge for software developers. Recently, different machine learning algorithms, such as GNNs, have been proposed to automate the pragma design via performance prediction. However, when applying the trained model on new kernels, the significant domain shift often leads to unsatisfactory performance. We propose a more domain-generalizable model structure: a two-level hierarchical Mixture of Experts (MoE), that can be flexibly adapted to any GNN model. Different expert networks can learn to deal with different regions in the representation space, and they can utilize similar patterns between the old kernels and new kernels. In the low-level MoE, we apply MoE on three natural granularities of a program: node, basic block, and graph. The high-level MoE learns to aggregate the three granularities for the final decision. To train the hierarchical MoE stably, we further propose a two-stage training method to avoid expert polarization. Extensive experiments verify the effectiveness of the proposed hierarchical MoE. We publicized our codes at https://github.com/weikai-li/HierarchicalMoE.

Yangchao Wu, Zongyue Qin, Alex Wong et al.

Speculative decoding is a technique to leverage hardware concurrency in order to enable multiple steps of token generation in a single forward pass, thus improving the efficiency of large-scale autoregressive (AR) Transformer models. State-space models (SSMs) are already more efficient than AR Transformers, since their state summarizes all past data with no need to cache or re-process tokens in the sliding window context. However, their state can also comprise thousands of tokens; so, speculative decoding has recently been extended to SSMs. Existing approaches, however, do not leverage the tree-based verification methods, since current SSMs lack the means to compute a token tree efficiently. We propose the first scalable algorithm to perform tree-based speculative decoding in state-space models (SSMs) and hybrid architectures of SSMs and Transformer layers. We exploit the structure of accumulated state transition matrices to facilitate tree-based speculative decoding with minimal overhead relative to current SSM implementations. Along with the algorithm, we describe a hardware-aware implementation that improves naive application of AR Transformer tree-based speculative decoding methods to SSMs. Furthermore, we outperform vanilla speculative decoding with SSMs even with a baseline drafting model and tree structure on three different benchmarks, opening up opportunities for further speed up with SSM and hybrid model inference. Code can be found at: https://github.com/wyc1997/stree.

Zifan He, Yingqi Cao, Zongyue Qin et al.

Transformer-based large language models (LLM) have been widely used in language processing applications. However, due to the memory constraints of the devices, most of them restrict the context window. Even though recurrent models in previous works can memorize past tokens to enable unlimited context and maintain effectiveness, they have ``flat'' memory architectures. Such architectures have limitations in selecting and filtering information. Since humans are good at learning and self-adjustment, we believe that imitating brain memory hierarchy is beneficial for model memorization. Thus, we propose the Hierarchical Memory Transformer (HMT), a novel framework that facilitates a model's long-context processing ability by imitating human memorization behavior. Leveraging memory-augmented segment-level recurrence, we organize the memory hierarchy by preserving tokens from early input segments, passing memory embeddings along the sequence, and recalling relevant information from history. Evaluating general language modeling, question-answering tasks, and the summarization task, we show that HMT consistently improves the long-context processing ability of existing models. Furthermore, HMT achieves a comparable or superior generation quality to long-context LLMs with $2 \sim 57\times$ fewer parameters and $2.5 \sim 116\times$ less inference memory, significantly outperforming previous memory-augmented models. Code on Github: https://github.com/OswaldHe/HMT-pytorch.

Zongyue Qin, Shichang Zhang, Mingxuan Ju et al.

Link prediction is a crucial graph-learning task with applications including citation prediction and product recommendation. Distilling Graph Neural Networks (GNNs) teachers into Multi-Layer Perceptrons (MLPs) students has emerged as an effective approach to achieve strong performance and reducing computational cost by removing graph dependency. However, existing distillation methods only use standard GNNs and overlook alternative teachers such as specialized model for link prediction (GNN4LP) and heuristic methods (e.g., common neighbors). This paper first explores the impact of different teachers in GNN-to-MLP distillation. Surprisingly, we find that stronger teachers do not always produce stronger students: MLPs distilled from GNN4LP can underperform those distilled from simpler GNNs, while weaker heuristic methods can teach MLPs to near-GNN performance with drastically reduced training costs. Building on these insights, we propose Ensemble Heuristic-Distilled MLPs (EHDM), which eliminates graph dependencies while effectively integrating complementary signals via a gating mechanism. Experiments on ten datasets show an average 7.93% improvement over previous GNN-to-MLP approaches with 1.95-3.32 times less training time, indicating EHDM is an efficient and effective link prediction method.

Zongyue Qin, Yunsheng Bai, Atefeh Sohrabizadeh et al.

In recent years, domain-specific accelerators (DSAs) have gained popularity for applications such as deep learning and autonomous driving. To facilitate DSA designs, programmers use high-level synthesis (HLS) to compile a high-level description written in C/C++ into a design with low-level hardware description languages that eventually synthesize DSAs on circuits. However, creating a high-quality HLS design still demands significant domain knowledge, particularly in microarchitecture decisions expressed as \textit{pragmas}. Thus, it is desirable to automate such decisions with the help of machine learning for predicting the quality of HLS designs, requiring a deeper understanding of the program that consists of original code and pragmas. Naturally, these programs can be considered as sequence data. In addition, these programs can be compiled and converted into a control data flow graph (CDFG). But existing works either fail to leverage both modalities or combine the two in shallow or coarse ways. We propose ProgSG, a model that allows interaction between the source code sequence modality and the graph modality in a deep and fine-grained way. To alleviate the scarcity of labeled designs, a pre-training method is proposed based on a suite of compiler's data flow analysis tasks. Experimental results show that ProgSG reduces the RMSE of design performance predictions by up to $22\%$, and identifies designs with an average of $1.10\times$ and $1.26\times$ (up to $8.17\times$ and $13.31\times$) performance improvement in design space exploration (DSE) task compared to HARP and AutoDSE, respectively.

Yunsheng Bai, Atefeh Sohrabizadeh, Zongyue Qin et al.

Recent years have witnessed the growing popularity of domain-specific accelerators (DSAs), such as Google's TPUs, for accelerating various applications such as deep learning, search, autonomous driving, etc. To facilitate DSA designs, high-level synthesis (HLS) is used, which allows a developer to compile a high-level description in the form of software code in C and C++ into a design in low-level hardware description languages (such as VHDL or Verilog) and eventually synthesized into a DSA on an ASIC (application-specific integrated circuit) or FPGA (field-programmable gate arrays). However, existing HLS tools still require microarchitecture decisions, expressed in terms of pragmas (such as directives for parallelization and pipelining). To enable more people to design DSAs, it is desirable to automate such decisions with the help of deep learning for predicting the quality of HLS designs. This requires us a deeper understanding of the program, which is a combination of original code and pragmas. Naturally, these programs can be considered as sequence data, for which large language models (LLM) can help. In addition, these programs can be compiled and converted into a control data flow graph (CDFG), and the compiler also provides fine-grained alignment between the code tokens and the CDFG nodes. However, existing works either fail to leverage both modalities or combine the two in shallow or coarse ways. We propose ProgSG allowing the source code sequence modality and the graph modalities to interact with each other in a deep and fine-grained way. To alleviate the scarcity of labeled designs, a pre-training method is proposed based on a suite of compiler's data flow analysis tasks. Experimental results on two benchmark datasets show the superiority of ProgSG over baseline methods that either only consider one modality or combine the two without utilizing the alignment information.