Ningyi Xu

Yijia Zhang, Sicheng Zhang, Shijie Cao et al.

Large language models (LLMs) show great performance in various tasks, but face deployment challenges from limited memory capacity and bandwidth. Low-bit weight quantization can save memory and accelerate inference. Although floating-point (FP) formats show good performance in LLM quantization, they tend to perform poorly with small group sizes or sub-4 bits. We find the reason is that the absence of asymmetry in previous FP quantization makes it unsuitable for handling asymmetric value distribution of LLM weight tensors. In this work, we propose asymmetric FP quantization (AFPQ), which sets separate scales for positive and negative values. Our method leads to large accuracy improvements and can be easily plugged into other quantization methods, including GPTQ and AWQ, for better performance. Besides, no additional storage is needed compared with asymmetric integer (INT) quantization. The code is available at https://github.com/zhangsichengsjtu/AFPQ.

Xiaoyu Zhang, Guangwei Liu, Zihao Liu et al.

In autonomous driving, there is growing interest in end-to-end online vectorized map perception in bird's-eye-view (BEV) space, with an expectation that it could replace traditional high-cost offline high-definition (HD) maps. However, the accuracy and robustness of these methods can be easily compromised in challenging conditions, such as occlusion or adverse weather, when relying only on onboard sensors. In this paper, we propose HRMapNet, leveraging a low-cost Historical Rasterized Map to enhance online vectorized map perception. The historical rasterized map can be easily constructed from past predicted vectorized results and provides valuable complementary information. To fully exploit a historical map, we propose two novel modules to enhance BEV features and map element queries. For BEV features, we employ a feature aggregation module to encode features from both onboard images and the historical map. For map element queries, we design a query initialization module to endow queries with priors from the historical map. The two modules contribute to leveraging map information in online perception. Our HRMapNet can be integrated with most online vectorized map perception methods. We integrate it in two state-of-the-art methods, significantly improving their performance on both the nuScenes and Argoverse 2 datasets. The source code is released at https://github.com/HXMap/HRMapNet.

Yongqiang Yao, Jingru Tan, Feizhao Zhang et al.

Vision-language instruction-tuning models have recently achieved significant performance improvements. In this work, we discover that large-scale 3D parallel training on those models leads to an imbalanced computation load across different devices. The vision and language parts are inherently heterogeneous: their data distribution and model architecture differ significantly, which affects distributed training efficiency. To address this issue, we rebalance the computational load from data, model, and memory perspectives, achieving more balanced computation across devices. Specifically, for the data, instances are grouped into new balanced mini-batches within and across devices. A search-based method is employed for the model to achieve a more balanced partitioning. For memory optimization, we adaptively adjust the re-computation strategy for each partition to utilize the available memory fully. These three perspectives are not independent but are closely connected, forming an omniverse balanced training framework. Extensive experiments are conducted to validate the effectiveness of our method. Compared with the open-source training code of InternVL-Chat, training time is reduced greatly, achieving about 1.8$\times$ speed-up. Our method's efficacy and generalizability are further validated across various models and datasets. Codes will be released at https://github.com/ModelTC/OmniBal.

Dayou Du, Yijia Zhang, Shijie Cao et al.

The upscaling of Large Language Models (LLMs) has yielded impressive advances in natural language processing, yet it also poses significant deployment challenges. Weight quantization has emerged as a widely embraced solution to reduce memory and computational demands. This paper introduces BitDistiller, a framework that synergizes Quantization-Aware Training (QAT) with Knowledge Distillation (KD) to boost the performance of LLMs at ultra-low precisions (sub-4-bit). Specifically, BitDistiller first incorporates a tailored asymmetric quantization and clipping technique to maximally preserve the fidelity of quantized weights, and then proposes a novel Confidence-Aware Kullback-Leibler Divergence (CAKLD) objective, which is employed in a self-distillation manner to enable faster convergence and superior model performance. Empirical evaluations demonstrate that BitDistiller significantly surpasses existing methods in both 3-bit and 2-bit configurations on general language understanding and complex reasoning benchmarks. Notably, BitDistiller is shown to be more cost-effective, demanding fewer data and training resources. The code is available at https://github.com/DD-DuDa/BitDistiller.

Jinhao Li, Shan Huang, Jiaming Xu et al.

We propose a Mamba accelerator with reconfigurable architecture, MARCA.We propose three novel approaches in this paper. (1) Reduction alternative PE array architecture for both linear and element-wise operations. For linear operations, the reduction tree connected to PE arrays is enabled and executes the reduction operation. For element-wise operations, the reduction tree is disabled and the output bypasses. (2) Reusable nonlinear function unit based on the reconfigurable PE. We decompose the exponential function into element-wise operations and a shift operation by a fast biased exponential algorithm, and the activation function (SiLU) into a range detection and element-wise operations by a piecewise approximation algorithm. Thus, the reconfigurable PEs are reused to execute nonlinear functions with negligible accuracy loss.(3) Intra-operation and inter-operation buffer management strategy. We propose intra-operation buffer management strategy to maximize input data sharing for linear operations within operations, and inter-operation strategy for element-wise operations between operations. We conduct extensive experiments on Mamba model families with different sizes.MARCA achieves up to 463.22$\times$/11.66$\times$ speedup and up to 9761.42$\times$/242.52$\times$ energy efficiency compared to Intel Xeon 8358P CPU and NVIDIA Tesla A100 GPU implementations, respectively.

Qiao Sun, Shiduo Zhang, Danjiao Ma et al.

Motion prediction and planning are vital tasks in autonomous driving, and recent efforts have shifted to machine learning-based approaches. The challenges include understanding diverse road topologies, reasoning traffic dynamics over a long time horizon, interpreting heterogeneous behaviors, and generating policies in a large continuous state space. Inspired by the success of large language models in addressing similar complexities through model scaling, we introduce a scalable trajectory model called State Transformer (STR). STR reformulates the motion prediction and motion planning problems by arranging observations, states, and actions into one unified sequence modeling task. Our approach unites trajectory generation problems with other sequence modeling problems, powering rapid iterations with breakthroughs in neighbor domains such as language modeling. Remarkably, experimental results reveal that large trajectory models (LTMs), such as STR, adhere to the scaling laws by presenting outstanding adaptability and learning efficiency. Qualitative results further demonstrate that LTMs are capable of making plausible predictions in scenarios that diverge significantly from the training data distribution. LTMs also learn to make complex reasonings for long-term planning, without explicit loss designs or costly high-level annotations.

Zihao Liu, Xiaoyu Zhang, Guangwei Liu et al.

In autonomous driving, the high-definition (HD) map plays a crucial role in localization and planning. Recently, several methods have facilitated end-to-end online map construction in DETR-like frameworks. However, little attention has been paid to the potential capabilities of exploring the query mechanism for map elements. This paper introduces MapQR, an end-to-end method with an emphasis on enhancing query capabilities for constructing online vectorized maps. To probe desirable information efficiently, MapQR utilizes a novel query design, called scatter-and-gather query, which is modelled by separate content and position parts explicitly. The base map instance queries are scattered to different reference points and added with positional embeddings to probe information from BEV features. Then these scatted queries are gathered back to enhance information within each map instance. Together with a simple and effective improvement of a BEV encoder, the proposed MapQR achieves the best mean average precision (mAP) and maintains good efficiency on both nuScenes and Argoverse 2. In addition, integrating our query design into other models can boost their performance significantly. The source code is available at https://github.com/HXMap/MapQR.

Ziyang Zheng, Shan Huang, Jianyuan Zhong et al.

Circuit representation learning has become pivotal in electronic design automation, enabling critical tasks such as testability analysis, logic reasoning, power estimation, and SAT solving. However, existing models face significant challenges in scaling to large circuits due to limitations like over-squashing in graph neural networks and the quadratic complexity of transformer-based models. To address these issues, we introduce DeepGate4, a scalable and efficient graph transformer specifically designed for large-scale circuits. DeepGate4 incorporates several key innovations: (1) an update strategy tailored for circuit graphs, which reduce memory complexity to sub-linear and is adaptable to any graph transformer; (2) a GAT-based sparse transformer with global and local structural encodings for AIGs; and (3) an inference acceleration CUDA kernel that fully exploit the unique sparsity patterns of AIGs. Our extensive experiments on the ITC99 and EPFL benchmarks show that DeepGate4 significantly surpasses state-of-the-art methods, achieving 15.5% and 31.1% performance improvements over the next-best models. Furthermore, the Fused-DeepGate4 variant reduces runtime by 35.1% and memory usage by 46.8%, making it highly efficient for large-scale circuit analysis. These results demonstrate the potential of DeepGate4 to handle complex EDA tasks while offering superior scalability and efficiency. Code is available at https://github.com/zyzheng17/DeepGate4-ICLR-25.

Yongqiang Yao, Jingru Tan, Kaihuan Liang et al.

Training Long-Context Large Language Models (LLMs) is challenging, as hybrid training with long-context and short-context data often leads to workload imbalances. Existing works mainly use data packing to alleviate this issue, but fail to consider imbalanced attention computation and wasted communication overhead. This paper proposes Hierarchical Balance Packing (HBP), which designs a novel batch-construction method and training recipe to address those inefficiencies. In particular, the HBP constructs multi-level data packing groups, each optimized with a distinct packing length. It assigns training samples to their optimal groups and configures each group with the most effective settings, including sequential parallelism degree and gradient checkpointing configuration. To effectively utilize multi-level groups of data, we design a dynamic training pipeline specifically tailored to HBP, including curriculum learning, adaptive sequential parallelism, and stable loss. Our extensive experiments demonstrate that our method significantly reduces training time over multiple datasets and open-source models while maintaining strong performance. For the largest DeepSeek-V2 (236B) MoE model, our method speeds up the training by 2.4$\times$ with competitive performance. Codes will be released at https://github.com/ModelTC/HBP.

Jinhao Li, Yuxuan Cong, Yingqiao Wang et al.

Diffusion policies have recently emerged as a powerful paradigm for visuomotor control in robotic manipulation due to their ability to model the distribution of action sequences and capture multimodality. However, iterative denoising leads to substantial inference latency, limiting control frequency in real-time closed-loop systems. Existing acceleration methods either reduce sampling steps, bypass diffusion through direct prediction, or reuse past actions, but often struggle to jointly preserve action quality and achieve consistently low latency. In this work, we propose STEP, a lightweight spatiotemporal consistency prediction mechanism to construct high-quality warm-start actions that are both distributionally close to the target action and temporally consistent, without compromising the generative capability of the original diffusion policy. Then, we propose a velocity-aware perturbation injection mechanism that adaptively modulates actuation excitation based on temporal action variation to prevent execution stall especially for real-world tasks. We further provide a theoretical analysis showing that the proposed prediction induces a locally contractive mapping, ensuring convergence of action errors during diffusion refinement. We conduct extensive evaluations on nine simulated benchmarks and two real-world tasks. Notably, STEP with 2 steps can achieve an average 21.6% and 27.5% higher success rate than BRIDGER and DDIM on the RoboMimic benchmark and real-world tasks, respectively. These results demonstrate that STEP consistently advances the Pareto frontier of inference latency and success rate over existing methods.

Jiancai Ye, Jun Liu, Qingchen Li et al.

Although Key-Value (KV) Cache is essential for efficient large language models (LLMs) inference, its growing memory footprint in long-context scenarios poses a significant bottleneck, making KVCache compression crucial. Current compression methods rely on rigid splitting strategies, such as fixed intervals or pre-defined delimiters. We observe that rigid splitting suffers from significant accuracy degradation (ranging from 5.5% to 55.1%) across different scenarios, owing to the scenario-dependent nature of the semantic boundaries. This highlights the necessity of dynamic semantic splitting to match semantics. To achieve this, we face two challenges. (1) Improper delimiter selection misaligns semantics with the KVCache, resulting in 28.6% accuracy loss. (2) Variable-length blocks after splitting introduce over 73.1% additional inference overhead. To address the above challenges, we propose DynSplit-KV, a KVCache compression method that dynamically identifies delimiters for splitting. We propose: (1) a dynamic importance-aware delimiter selection strategy, improving accuracy by 49.9%. (2) A uniform mapping strategy that transforms variable-length semantic blocks into a fixed-length format, reducing inference overhead by 4.9x. Experiments show that DynSplit-KV achieves the highest accuracy, 2.2x speedup compared with FlashAttention and 2.6x peak memory reduction in long-context scenarios.

Derun Li, Changye Li, Yue Wang et al.

Generating human-like and adaptive trajectories is essential for autonomous driving in dynamic environments. While generative models have shown promise in synthesizing feasible trajectories, they often fail to capture the nuanced variability of personalized driving styles due to dataset biases and distributional shifts. To address this, we introduce TrajHF, a human feedback-driven finetuning framework for generative trajectory models, designed to align motion planning with diverse driving styles. TrajHF incorporates multi-conditional denoiser and reinforcement learning with human feedback to refine multi-modal trajectory generation beyond conventional imitation learning. This enables better alignment with human driving preferences while maintaining safety and feasibility constraints. TrajHF achieves performance comparable to the state-of-the-art on NavSim benchmark. TrajHF sets a new paradigm for personalized and adaptable trajectory generation in autonomous driving.

Yijia Zhang, Zhihong Gou, Shijie Cao et al.

Deep Neural Networks (DNNs) have revolutionized various fields, but their deployment on GPUs often leads to significant energy consumption. Unlike existing methods for reducing GPU energy consumption, which are either hardware-inflexible or limited by workload constraints, this paper addresses the problem at the GPU kernel level. We propose a novel search-based compilation method to generate energy-efficient GPU kernels by incorporating energy efficiency into the search process. To accelerate the energy evaluation process, we develop an accurate energy cost model based on high-level kernel features. Furthermore, we introduce a dynamic updating strategy for the energy cost model, reducing the need for on-device energy measurements and accelerating the search process. Our evaluation demonstrates that the proposed approach can generate GPU kernels with up to 21.69% reduced energy consumption while maintaining low latency.

Changran Xu, Yi Liu, Yunhao Zhou et al.

The rapid advancement of large language models (LLMs) has revolutionized code generation tasks across various programming languages. However, the unique characteristics of programming languages, particularly those like Verilog with specific syntax and lower representation in training datasets, pose significant challenges for conventional tokenization and decoding approaches. In this paper, we introduce a novel application of speculative decoding for Verilog code generation, showing that it can improve both inference speed and output quality, effectively achieving speed and quality all in one. Unlike standard LLM tokenization schemes, which often fragment meaningful code structures, our approach aligns decoding stops with syntactically significant tokens, making it easier for models to learn the token distribution. This refinement addresses inherent tokenization issues and enhances the model's ability to capture Verilog's logical constructs more effectively. Our experimental results show that our method achieves up to a 5.05x speedup in Verilog code generation and increases pass@10 functional accuracy on RTLLM by up to 17.19% compared to conventional training strategies. These findings highlight speculative decoding as a promising approach to bridge the quality gap in code generation for specialized programming languages.

Wenqiang Wang, Yijia Zhang, Zikai Zhang et al.

As large language models (LLMs) demonstrate powerful capabilities, deploying them on edge devices has become increasingly crucial, offering advantages in privacy and real-time interaction. QLoRA has emerged as the standard approach for on-device LLMs, leveraging quantized models to reduce memory and computational costs while utilizing LoRA for task-specific adaptability. In this work, we propose ROMA, a QLoRA accelerator with a hybrid storage architecture that uses ROM for quantized base models and SRAM for LoRA weights and KV cache. Our insight is that the quantized base model is stable and converged, making it well-suited for ROM storage. Meanwhile, LoRA modules offer the flexibility to adapt to new data without requiring updates to the base model. To further reduce the area cost of ROM, we introduce a novel B-ROM design and integrate it with the compute unit to form a fused cell for efficient use of chip resources. ROMA can effectively store both a 4-bit 3B and a 2-bit 8B LLaMA model entirely on-chip, achieving a notable generation speed exceeding 20,000 tokens/s without requiring external memory.

Yijia Zhang, Yibo Han, Shijie Cao et al.

Running out of GPU memory has become a main bottleneck for large-scale DNN training. How to reduce the memory footprint during training has received intensive research attention. We find that previous gradient accumulation reduces activation memory but fails to be compatible with gradient memory reduction due to a contradiction between preserving gradients and releasing gradients. To address this issue, we propose a novel optimizer accumulation method for Adam, named Adam Accumulation (AdamA), which enables reducing both activation and gradient memory. Specifically, AdamA directly integrates gradients into optimizer states and accumulates optimizer states over micro-batches, so that gradients can be released immediately after use. We mathematically and experimentally demonstrate AdamA yields the same convergence properties as Adam. Evaluated on transformer-based models, AdamA achieves up to 23% memory reduction compared to gradient accumulation with less than 2% degradation in training throughput. Notably, AdamA can work together with memory reduction methods for optimizer states to fit 1.26x~3.14x larger models over PyTorch and DeepSpeed baseline on GPUs with different memory capacities.

Yijia Zhang, Lingran Zhao, Shijie Cao et al.

Efficient deployment of large language models (LLMs) necessitates low-bit quantization to minimize model size and inference cost. While low-bit integer formats (e.g., INT8/INT4) have been the conventional choice, emerging low-bit floating-point formats (e.g., FP8/FP4) offer a compelling alternative and are gaining support from cutting-edge hardware, such as NVIDIA's H100 GPU. However, the superiority of low-bit INT versus FP formats for quantization on LLMs remains unclear. In this study, we conduct a comparative analysis of INT and FP quantization with the same bit-width, revealing that the optimal quantization format varies across different layers due to the complexity and diversity of tensor distribution. Consequently, we advocate the Mixture of Formats Quantization (MoFQ), which selects the optimal format on a layer-wise basis. This simple yet effective approach achieves state-of-the-art results in both weight-only (W-only) and weight-activation (WA) post-training quantization scenarios when tested on LLaMA across various tasks. In 4-bit W-only quantization, MoFQ surpasses GPTQ without complex hyperparameter tuning and with an order of magnitude faster quantization speed. While in 8-bit WA quantization, MoFQ significantly outperforms INT/FP-only methods, achieving performance close to the full precision model. Notably, MoFQ incurs no hardware overhead compared to INT/FP-only quantization, as the bit-width remains unchanged.