Search

Large Language Models (LLMs) have been widely adopted to process long-context tasks. However, the large memory overhead of the key-value (KV) cache poses significant challenges in long-context scenarios. Existing training-free KV cache compression methods typically focus on quantization and token pruning, which have compression limits, and excessive sparsity can lead to severe performance degradation. Other methods design new architectures with less KV overhead but require significant training overhead. To address the above two drawbacks, we further explore the redundancy in the channel dimension and apply an architecture-level design with minor training costs. Therefore, we introduce CSKV, a training-efficient Channel Shrinking technique for KV cache compression: (1) We first analyze the singular value distribution of the KV cache, revealing significant redundancy and compression potential along the channel dimension. Based on this observation, we propose using low-rank decomposition for key and value layers and storing the low-dimension features. (2) To preserve model performance, we introduce a bi-branch KV cache, including a window-based full-precision KV cache and a low-precision compressed KV cache. (3) To reduce the training costs, we minimize the layer-wise reconstruction loss for the compressed KV cache instead of retraining the entire LLMs. Extensive experiments show that CSKV can reduce the memory overhead of the KV cache by 80% while maintaining the model's long-context capability. Moreover, we show that our method can be seamlessly combined with quantization to further reduce the memory overhead, achieving a compression ratio of up to 95%.

This paper introduces GateAttentionPose, an innovative approach that enhances the UniRepLKNet architecture for pose estimation tasks. We present two key contributions: the Agent Attention module and the Gate-Enhanced Feedforward Block (GEFB). The Agent Attention module replaces large kernel convolutions, significantly improving computational efficiency while preserving global context modeling. The GEFB augments feature extraction and processing capabilities, particularly in complex scenes. Extensive evaluations on COCO and MPII datasets demonstrate that GateAttentionPose outperforms existing state-of-the-art methods, including the original UniRepLKNet, achieving superior or comparable results with improved efficiency. Our approach offers a robust solution for pose estimation across diverse applications, including autonomous driving, human motion capture, and virtual reality.

Sparse attention can effectively mitigate the significant memory and throughput demands of Large Language Models (LLMs) in long contexts. Existing methods typically employ a uniform sparse attention mask, applying the same sparse pattern across different attention heads and input lengths. However, this uniform approach fails to capture the diverse attention patterns inherent in LLMs, ignoring their distinct accuracy-latency trade-offs. To address this challenge, we propose the Mixture of Attention (MoA), which automatically tailors distinct sparse attention configurations to different heads and layers. MoA constructs and navigates a search space of various attention patterns and their scaling rules relative to input sequence lengths. It profiles the model, evaluates potential configurations, and pinpoints the optimal sparse attention compression plan. MoA adapts to varying input sizes, revealing that some attention heads expand their focus to accommodate longer sequences, while other heads consistently concentrate on fixed-length local contexts. Experiments show that MoA increases the effective context length by $3.9\times$ with the same average attention span, boosting retrieval accuracy by $1.5-7.1\times$ over the uniform-attention baseline across Vicuna-7B, Vicuna-13B, and Llama3-8B models. Moreover, MoA narrows the capability gaps between sparse and dense models, reducing the maximum relative performance drop from $9\%-36\%$ to within $5\%$ across two long-context understanding benchmarks. MoA achieves a $1.2-1.4\times$ GPU memory reduction and boosts decode throughput by $5.5-6.7 \times$ for 7B and 13B dense models on a single GPU, with minimal impact on performance., Comment: 10 pages

Teaching to improve student models (e.g., knowledge distillation) is an extensively studied methodology in LLMs. However, for humans, teaching not only improves students but also improves teachers. We ask: Can LLMs also learn by teaching (LbT)? If yes, we can potentially unlock the possibility of continuously advancing the models without solely relying on human-produced data or stronger models. In this paper, we provide a preliminary exploration of this ambitious agenda. We show that LbT ideas can be incorporated into existing LLM training/prompting pipelines and provide noticeable improvements. Specifically, we design three methods, each mimicking one of the three levels of LbT in humans: observing students' feedback, learning from the feedback, and learning iteratively, with the goals of improving answer accuracy without training and improving models' inherent capability with fine-tuning. The findings are encouraging. For example, similar to LbT in human, we see that: (1) LbT can induce weak-to-strong generalization: strong models can improve themselves by teaching other weak models; (2) Diversity in students might help: teaching multiple students could be better than teaching one student or the teacher itself. We hope that this early promise can inspire future research on LbT and more broadly adopting the advanced techniques in education to improve LLMs. The code is available at https://github.com/imagination-research/lbt., Comment: Under review

Diffusion transformers (DiTs) have exhibited remarkable performance in visual generation tasks, such as generating realistic images or videos based on textual instructions. However, larger model sizes and multi-frame processing for video generation lead to increased computational and memory costs, posing challenges for practical deployment on edge devices. Post-Training Quantization (PTQ) is an effective method for reducing memory costs and computational complexity. When quantizing diffusion transformers, we find that applying existing diffusion quantization methods designed for U-Net faces challenges in preserving quality. After analyzing the major challenges for quantizing diffusion transformers, we design an improved quantization scheme: "ViDiT-Q": Video and Image Diffusion Transformer Quantization) to address these issues. Furthermore, we identify highly sensitive layers and timesteps hinder quantization for lower bit-widths. To tackle this, we improve ViDiT-Q with a novel metric-decoupled mixed-precision quantization method (ViDiT-Q-MP). We validate the effectiveness of ViDiT-Q across a variety of text-to-image and video models. While baseline quantization methods fail at W8A8 and produce unreadable content at W4A8, ViDiT-Q achieves lossless W8A8 quantization. ViDiTQ-MP achieves W4A8 with negligible visual quality degradation, resulting in a 2.5x memory optimization and a 1.5x latency speedup., Comment: Project Page: https://a-suozhang.xyz/viditq.github.io/

Large Language Models (LLMs) have attracted extensive attention due to their remarkable performance across various tasks. However, the substantial computational and memory requirements of LLM inference pose challenges for deployment in resource-constrained scenarios. Efforts within the field have been directed towards developing techniques aimed at enhancing the efficiency of LLM inference. This paper presents a comprehensive survey of the existing literature on efficient LLM inference. We start by analyzing the primary causes of the inefficient LLM inference, i.e., the large model size, the quadratic-complexity attention operation, and the auto-regressive decoding approach. Then, we introduce a comprehensive taxonomy that organizes the current literature into data-level, model-level, and system-level optimization. Moreover, the paper includes comparative experiments on representative methods within critical sub-fields to provide quantitative insights. Last but not least, we provide some knowledge summary and discuss future research directions.

Post-training quantization (PTQ) has emerged as a promising technique to reduce the cost of large language models (LLMs). Specifically, PTQ can effectively mitigate memory consumption and reduce computational overhead in LLMs. To meet the requirements of both high efficiency and performance across diverse scenarios, a comprehensive evaluation of quantized LLMs is essential to guide the selection of quantization methods. This paper presents a thorough evaluation of these factors by evaluating the effect of PTQ on Weight, Activation, and KV Cache on 11 model families, including OPT, LLaMA2, Falcon, Bloomz, Mistral, ChatGLM, Vicuna, LongChat, StableLM, Gemma, and Mamba, with parameters ranging from 125M to 180B. The evaluation encompasses five types of tasks: basic NLP, emergent ability, trustworthiness, dialogue, and long-context tasks. Moreover, we also evaluate the state-of-the-art (SOTA) quantization methods to demonstrate their applicability. Based on the extensive experiments, we systematically summarize the effect of quantization, provide recommendations to apply quantization techniques, and point out future directions. The code can be found in https://github.com/thu-nics/qllm-eval.

State-of-the-art large language models (LLMs) are now claiming remarkable supported context lengths of 256k or even more. In contrast, the average context lengths of mainstream benchmarks are insufficient (5k-21k), and they suffer from potential knowledge leakage and inaccurate metrics, resulting in biased evaluation. This paper introduces LV-Eval, a challenging long-context benchmark with five length levels (16k, 32k, 64k, 128k, and 256k) reaching up to 256k words. LV-Eval features two main tasks, single-hop QA and multi-hop QA, comprising 11 bilingual datasets. The design of LV-Eval has incorporated three key techniques, namely confusing facts insertion, keyword and phrase replacement, and keyword-recall-based metric design. The advantages of LV-Eval include controllable evaluation across different context lengths, challenging test instances with confusing facts, mitigated knowledge leakage, and more objective evaluations. We evaluate 10 LLMs on LV-Eval and conduct ablation studies on the techniques used in LV-Eval construction. The results reveal that: (i) Commercial LLMs generally outperform open-source LLMs when evaluated within length levels shorter than their claimed context length. However, their overall performance is surpassed by open-source LLMs with longer context lengths. (ii) Extremely long-context LLMs, such as Yi-6B-200k, exhibit a relatively gentle degradation of performance, but their absolute performances may not necessarily be higher than those of LLMs with shorter context lengths. (iii) LLMs' performances can significantly degrade in the presence of confusing information, especially in the pressure test of "needle in a haystack". (iv) Issues related to knowledge leakage and inaccurate metrics introduce bias in evaluation, and these concerns are alleviated in LV-Eval. All datasets and evaluation codes are released at: https://github.com/infinigence/LVEval.

Transformer-based Large Language Models (LLMs) have made a significant impact on various domains. However, LLMs' efficiency suffers from both heavy computation and memory overheads. Compression techniques like sparsification and quantization are commonly used to mitigate the gap between LLM's computation/memory overheads and hardware capacity. However, existing GPU and transformer-based accelerators cannot efficiently process compressed LLMs, due to the following unresolved challenges: low computational efficiency, underutilized memory bandwidth, and large compilation overheads. This paper proposes FlightLLM, enabling efficient LLMs inference with a complete mapping flow on FPGAs. In FlightLLM, we highlight an innovative solution that the computation and memory overhead of LLMs can be solved by utilizing FPGA-specific resources (e.g., DSP48 and heterogeneous memory hierarchy). We propose a configurable sparse DSP chain to support different sparsity patterns with high computation efficiency. Second, we propose an always-on-chip decode scheme to boost memory bandwidth with mixed-precision support. Finally, to make FlightLLM available for real-world LLMs, we propose a length adaptive compilation method to reduce the compilation overhead. Implemented on the Xilinx Alveo U280 FPGA, FlightLLM achieves 6.0$\times$ higher energy efficiency and 1.8$\times$ better cost efficiency against commercial GPUs (e.g., NVIDIA V100S) on modern LLMs (e.g., LLaMA2-7B) using vLLM and SmoothQuant under the batch size of one. FlightLLM beats NVIDIA A100 GPU with 1.2$\times$ higher throughput using the latest Versal VHK158 FPGA., Comment: Accepted to FPGA'24

Large language models (LLMs) have demonstrated impressive abilities in various domains while the inference cost is expensive. Many previous studies exploit quantization methods to reduce LLM inference cost by reducing latency and memory consumption. Applying 2-bit single-precision weight quantization brings >3% accuracy loss, so the state-of-the-art methods use mixed-precision methods for LLMs (e.g. Llama2-7b, etc.) to improve the accuracy. However, challenges still exist: (1) Uneven distribution in weight matrix. (2) Large speed degradation by adding sparse outliers. (3) Time-consuming dequantization operations on GPUs. To tackle these challenges and enable fast and efficient LLM inference on GPUs, we propose the following techniques in this paper. (1) Intra-weight mixed-precision quantization. (2) Exclusive 2-bit sparse outlier with minimum speed degradation. (3) Asynchronous dequantization. We conduct extensive experiments on different model families (e.g. Llama3, etc.) and model sizes. We achieve 2.91-bit for each weight considering all scales/zeros for different models with negligible loss. As a result, with our 2/4/16 mixed-precision quantization for each weight matrix and asynchronous dequantization during inference, our design achieves an end-to-end speedup for Llama2-7b is 1.74x over the original model, and we reduce both runtime cost and total cost by up to 2.53x and 2.29x with less GPU requirements.

This paper reports initial results from a usability study conducted in the formative and user-centered design phase of a larger project to translate an existing, science-focused educational technology for neurotypical middle school students into a new, social-reasoning-focused educational technology for students on the autism spectrum. Participants in our study included both adolescents on the autism spectrum and typically developing adolescents, who were asked to complete the Betty's Brain educational-technology-based science activity as well as a social-reasoning movie question-answering activity. Results include qualitative observations of general student engagement and challenges as well as quantitative measures of performance and eye gaze, including key differences observed across our two sample groups, with the goal of informing the design and adaptation of future technology-based interventions. Our findings suggest specific considerations for designing educational technologies for adolescents on the autism spectrum, including: (1) finding ways to help students follow instructional/tutorial portions of new technologies, especially when lengthy instructions and/or complex interfaces are involved; (2) proactively anticipating and finding ways to mitigate potential student episodes of frustration / dysregulation while using the technology; and (3) capitalizing on features of the technology found to be engaging/motivating for students. [This paper was published in: "Proceedings of the 42nd Annual Meeting of the Cognitive Science Society," Cognitive Science Society, 2020, pp. 1922-1928.]

Modern social intelligence includes the ability to watch videos and answer questions about social and theory-of-mind-related content, e.g., for a scene in Harry Potter, "Is the father really upset about the boys flying the car?" Social visual question answering (social VQA) is emerging as a valuable methodology for studying social reasoning in both humans (e.g., children with autism) and AI agents. However, this problem space spans enormous variations in both videos and questions. We discuss methods for creating and characterizing social VQA datasets, including 1) crowdsourcing versus in-house authoring, including sample comparisons of two new datasets that we created (TinySocial-Crowd and TinySocial-InHouse) and the previously existing Social-IQ dataset; 2) a new rubric for characterizing the difficulty and content of a given video; and 3) a new rubric for characterizing question types. We close by describing how having well-characterized social VQA datasets will enhance the explainability of AI agents and can also inform assessments and educational interventions for people., Comment: To appear in the Joint IEEE International Conference on Development and Learning and on Epigenetic Robotics (ICDL), 2020

A novel method for constructing robust and high-order accurate weighted essentially non-oscillatory (WENO) scheme is proposed in this paper. The method is mainly based on the WENO-Z type scheme, in which, an eighth-order global smoothness indicator (the square of the approximation of the fourth-order derivative on the five-point stencil used by the fifth-order WENO scheme) is used, and in order to keep the ENO property and robustness, the constant 1 used to calculate the un-normalized weights is replaced by a function of local smoothness indicators of candidate sub-stencils. This function is designed to have following adaptive property: if the five-point stencil contains a discontinuity, then the function approaches to a small value, otherwise, it approaches to a large value. Analysis and numerical results show that the resulted WENO-Z type (WENO-ZN) scheme is robust for capturing shock waves and, in smooth regions, achieves fifth-order accuracy at first-order critical point and fourth-order accuracy at second-order critical point.

High fidelity numerical simulation of compressible flow requires the numerical method being used to have both stable shock-capturing capability and high spectral resolution. Recently, a family of Targeted Essentially Non-Oscillatory (TENO) schemes are developed to fulfill such requirements. Although TENO has very low dissipation for smooth flow, it introduces a cutoff value $C_T$ to maintain the non-oscillatory shock-capturing property. $C_T$ is problem depended and therefore needs adjustments by trial and error for different problems. As $C_T$ actually controls the dissipation property of TENO, the choice of $C_T$ for better shock-capturing capability always means higher dissipation. To overcome this, in this paper, a local adaptive method is proposed for the choice of $C_T$. By introducing a novel adaptive function based on the WENO smoothness indicators, $C_T$ is dynamically adjusted form $1.0 \times 10.0^{-10}$ for lower dissipation to $1.0 \times 10.0^{-4}$ for better capturing of shock according to the smoothness of the reconstruction stencil. Numerical results of the new method are compared with those of the original TENO method and the TENO-A method (Fu et al., JCP, 2017). It reveals that the new method is capable of better suppressing numerical oscillations near discontinuities while maintaining the low dissipation property of TENO at lower extra computational cost.