Xianzhi Yu

Huiling Zhen, Weizhe Lin, Renxi Liu et al.

Diffusion large language models (DLLMs) have emerged as an alternative to autoregressive (AR) decoding with appealing efficiency and modeling properties, yet their implications for agentic multi-step decision making remain underexplored. We ask a concrete question: when the generation paradigm is changed but the agent framework and supervision are held fixed, do diffusion backbones induce systematically different planning and tool-use behaviors, and do these differences translate into end-to-end efficiency gains? We study this in a controlled setting by instantiating DLLM and AR backbones within the same agent workflow (DeepDiver) and performing matched agent-oriented fine-tuning on the same trajectory data, yielding diffusion-backed DLLM Agents and directly comparable AR agents. Across benchmarks and case studies, we find that, at comparable accuracy, DLLM Agents are on average over 30% faster end to end than AR agents, with some cases exceeding 8x speedup. Conditioned on correct task completion, DLLM Agents also require fewer interaction rounds and tool invocations, consistent with higher planner hit rates that converge earlier to a correct action path with less backtracking. We further identify two practical considerations for deploying diffusion backbones in tool-using agents. First, naive DLLM policies are more prone to structured tool-call failures, necessitating stronger tool-call-specific training to emit valid schemas and arguments. Second, for multi-turn inputs interleaving context and action spans, diffusion-style span corruption requires aligned attention masking to avoid spurious context-action information flow; without such alignment, performance degrades. Finally, we analyze attention dynamics across workflow stages and observe paradigm-specific coordination patterns, suggesting stronger global planning signals in diffusion-backed agents.

Zehua Pei, Hui-Ling Zhen, Xianzhi Yu et al.

Large language models can now process increasingly long inputs, yet their ability to effectively use information spread across long contexts remains limited. We trace this gap to how attention budget is spent during supervised fine-tuning (SFT) on long sequences: positional biases and attention sinks cause the model to allocate most of its attention to positionally privileged tokens rather than semantically relevant content. This training-time attention dilution (the starvation of content tokens in the attention distribution) weakens the gradient signal, limiting the model's ability to learn robust long-context capabilities. We introduce FocuSFT, a bilevel optimization framework that addresses this problem at training time. An inner loop adapts lightweight fast-weight parameters on the training context to form a parametric memory that concentrates attention on relevant content, and the outer loop performs SFT conditioned on this sharpened representation. Both loops apply bidirectional attention over context tokens while preserving causal masking for responses, reducing the causal asymmetry that gives rise to attention sinks and aligning inner-outer behavior. On BABILong, FocuSFT improves accuracy by up to +14pp across 4K--32K context lengths; on RULER, it raises CWE aggregation from 72.9\% to 81.1\% at 16K; and on GPQA with agentic tool use, it yields a 24\% relative gain in pass@1. Attention analysis shows that FocuSFT reduces attention sink mass by 529$\times$ and triples context engagement during training. Code: https://github.com/JarvisPei/FocuSFT

Xiaohao Liu, Xiaobo Xia, Manyi Zhang et al.

Quantization is pivotal for mitigating the significant memory and computational overhead of Large Language Models (LLMs). While emerging transformation-based methods have successfully enhanced quantization by projecting feature spaces onto smoother manifolds using orthogonal matrices, they typically enforce a rigid one-to-one transformation constraint. This static approach fails to account for the dynamic patterns inherent in input activations, particularly within diffusion LLMs (dLLMs) and Multimodal LLMs (MLLMs), where varying token types exhibit distinct distributions. To advance this, we propose FreeAct, a novel quantization framework that relaxes the static one-to-one constraint to accommodate dynamic activation disparities. Theoretically, we leverage the rank-deficient nature of activations to derive a solution space that extends beyond simple inverse matrices, enabling the decoupling of activation transformations from weights. Methodologically, FreeAct identifies token-specific dynamics (i.e., vision v.s. text, or masked tokens) and allocates distinct transformation matrices to the activation side, while maintaining a unified, static transformation for the weights. Extensive experiments across dLLMs and MLLMs demonstrate that FreeAct significantly outperforms baselines, up to 5.3% performance improvement, with in-depth analyses. Our code will be publicly released.

Yuxuan Sun, Ruikang Liu, Haoli Bai et al.

Recently, quantization has been widely used for the compression and acceleration of large language models (LLMs). Due to the outliers in LLMs, it is crucial to flatten weights and activations to minimize quantization error with equally spaced quantization points. Prior research explores various pre-quantization transformations to suppress outliers, such as per-channel scaling and Hadamard transformation. However, we observe that these transformed weights and activations can still exhibit steep and dispersed distributions. In this paper, we propose FlatQuant (Fast and Learnable Affine Transformation), a new post-training quantization approach that enhances the flatness of weights and activations. Our approach identifies optimal affine transformations for each linear layer, calibrated in hours via a lightweight objective. To reduce runtime overhead of affine transformation, we apply Kronecker product with two lightweight matrices, and fuse all operations in FlatQuant into a single kernel. Extensive experiments demonstrate that FlatQuant establishes a new state-of-the-art benchmark for quantization. For example, it achieves less than 1\% accuracy drop for W4A4 quantization on the LLaMA-3-70B model, surpassing SpinQuant by 7.5\%. Additionally, it provides up to 2.3x prefill speedup and 1.7x decoding speedup compared to the FP16 model. Code is available at: https://github.com/ruikangliu/FlatQuant.

Ruikang Liu, Yuxuan Sun, Manyi Zhang et al.

Recent advancements in reasoning language models have demonstrated remarkable performance in complex tasks, but their extended chain-of-thought reasoning process increases inference overhead. While quantization has been widely adopted to reduce the inference cost of large language models, its impact on reasoning models remains understudied. In this paper, we conduct the first systematic study on quantized reasoning models, evaluating the open-sourced DeepSeek-R1-Distilled Qwen and LLaMA families ranging from 1.5B to 70B parameters, QwQ-32B, and Qwen3-8B. Our investigation covers weight, KV cache, and activation quantization using state-of-the-art algorithms at varying bit-widths, with extensive evaluation across mathematical (AIME, MATH-500), scientific (GPQA), and programming (LiveCodeBench) reasoning benchmarks. Our findings reveal that while lossless quantization can be achieved with W8A8 or W4A16 quantization, lower bit-widths introduce significant accuracy risks. We further identify model size, model origin, and task difficulty as critical determinants of performance. Contrary to expectations, quantized models do not exhibit increased output lengths. In addition, strategically scaling the model sizes or reasoning steps can effectively enhance the performance. All quantized models and codes are open-sourced in https://github.com/ruikangliu/Quantized-Reasoning-Models.

Manyi Zhang, Ji-Fu Li, Zhongao Sun et al.

Autonomous agent systems such as OpenClaw introduce significant efficiency challenges due to long-context inputs and multi-turn reasoning. This results in prohibitively high computational and monetary costs in real-world development. While quantization is a standard approach for reducing cost and latency, its impact on agent performance in realistic scenarios remains unclear. In this work, we analyze quantization sensitivity across diverse complex workflows over OpenClaw, and show that precision requirements are highly task-dependent. Based on this observation, we propose QuantClaw, a plug-and-play precision routing plugin that dynamically assigns precision according to task characteristics. QuantClaw routes lightweight tasks to lower-cost configurations while preserving higher precision for demanding workloads, saving cost and accelerating inference without increasing user complexity. Experiments show that our QuantClaw maintains or improves task performance while reducing both latency and computational cost. Across a range of agent tasks, it achieves up to 21.4% cost savings and 15.7% latency reduction on GLM-5 (FP8 baseline). These results highlight the benefit of treating precision as a dynamic resource in agent systems.

Keyu Lv, Manyi Zhang, Xiaobo Xia et al.

Reasoning models excel at complex tasks such as coding and mathematics, yet their inference is often slow and token-inefficient. To improve the inference efficiency, post-training quantization (PTQ) usually comes with the cost of large accuracy drops, especially for reasoning tasks under low-bit settings. In this study, we present a systematic empirical study of quantization-aware training (QAT) for reasoning models. Our key findings include: (1) Knowledge distillation is a robust objective for reasoning models trained via either supervised fine-tuning or reinforcement learning; (2) PTQ provides a strong initialization for QAT, improving accuracy while reducing training cost; (3) Reinforcement learning remains feasible for quantized models given a viable cold start and yields additional gains; and (4) Aligning the PTQ calibration domain with the QAT training domain accelerates convergence and often improves the final accuracy. Finally, we consolidate these findings into an optimized workflow (Reasoning-QAT), and show that it consistently outperforms state-of-the-art PTQ methods across multiple LLM backbones and reasoning datasets. For instance, on Qwen3-0.6B, it surpasses GPTQ by 44.53% on MATH-500 and consistently recovers performance in the 2-bit regime.

Manyi Zhang, Ji-Fu Li, Zhongao Sun et al.

Microscaling Floating-Point (MXFP) has emerged as a promising low-precision format for large language models (LLMs). Despite various post-training quantization (PTQ) algorithms being proposed, they mostly focus on integer quantization, while their applicability and behavior under MXFP formats remain largely unexplored. To address this gap, this work conducts a systematic investigation of PTQ under MXFP formats, encompassing over 7 PTQ algorithms, 15 evaluation benchmarks, and 3 LLM families. The key findings include: 1) MXFP8 consistently achieves near-lossless performance, while MXFP4 introduces substantial accuracy degradation and remains challenging; 2) PTQ effectiveness under MXFP depends strongly on format compatibility, with some algorithmic paradigms being consistently more effective than others; 3) PTQ performance exhibits highly consistent trends across model families and modalities, in particular, quantization sensitivity is dominated by the language model rather than the vision encoder in multimodal LLMs; 4) The scaling factor of quantization is a critical error source in MXFP4, and a simple pre-scale optimization strategy can significantly mitigate its impact. Together, these results provide practical guidance on adapting existing PTQ methods to MXFP quantization.

Shengyin Sun, Yiming Li, Renxi Liu et al.

Diffusion large language models (dLLMs) generate text by iteratively denoising masked token sequences. Although dLLMs can predict all masked positions in parallel within each step, the large number of denoising iterations still makes inference expensive. This cost can be reduced spatially by unmasking multiple tokens per step, or temporally by collapsing multiple denoising steps into one verification call. We propose Parallel Speculative Decoding (PSD), a training-free framework that jointly improves inference along both axes. Using the confidence scores from a single forward pass, PSD selects positions to unmask via a configurable, adaptive unmasking policy and constructs multi-depth speculative drafts without extra model calls. A final batched verification pass then applies hierarchical acceptance, keeping the deepest draft that remains consistent with the updated predictions. Experiments on three dLLMs across reasoning and code generation tasks show that PSD achieves favorable trade-offs between inference efficiency and generation quality, reaching up to $5.5\times$ tokens per forward pass with accuracy comparable to greedy decoding.

Ji-Fu Li, Manyi Zhang, Xiaobo Xia et al.

Microscaling floating-point (MXFP) formats have emerged as a promising standard for deploying Multi-modal Large Language Models (MLLMs) and Large Language Models (LLMs) on modern accelerator architectures. However, existing Post-Training Quantization (PTQ) methods, particularly rotation-based techniques designed for integer formats, suffer from severe performance collapse when applied to MXFP4. Recent studies attribute this failure to a fundamental format mismatch: global orthogonal rotations inadvertently transfer outlier energy across quantization blocks, inducing new outliers that disrupt local block-wise scaling, while often creating bimodal activation distributions that underutilize the limited quantization range. To address these issues, we propose BATQuant (Block-wise Affine Transformation), which restricts transformations to align with MXFP granularity to prevent cross-block outlier propagation, while relaxing orthogonality constraints to optimize distribution shaping. To ensure parameter efficiency, we introduce Global and Private Kronecker (GPK) decomposition to effectively reduces storage and runtime overhead and incorporate Block-wise Learnable Clipping to suppress residual outliers. Extensive experiments on both MLLMs and LLMs demonstrate that BATQuant establishes new state-of-the-art results under aggressive W4A4KV16 configurations, recovering up to 96.43% of full-precision performance on multimodal benchmarks and clearly outperforming existing methods across diverse tasks.

Pengxiang Zhao, Hui-Ling Zhen, Xing Li et al.

As LLMs scale, low-bit floating-point formats like MXFP and NVFP4 offer new opportunities for precision and efficiency. In this work, we evaluate HiFloat (HiF8 and HiF4), a family of formats tailored for Ascend NPUs. Through rigorous comparison across weight-activation and KV-cache tasks, we provide three key insights: (1) INT8 suits narrow-range data, while floating-point formats excel with high-variance data; (2) in 4-bit regimes, HiF4's hierarchical scaling prevents the accuracy collapse seen in integer formats; and (3) HiFloat is fully compatible with state-of-the-art post-training quantization frameworks. Overall, HiFloat provides a solution for high-efficiency LLM inference on NPUs.

Xinrui Chen, Haoli Bai, Tao Yuan et al.

Layer pruning has emerged as a widely used technique for compressing large language models (LLMs). However, existing layer pruning approaches often incur substantial performance degradation. We identify the majority of this degradation to a single yet previously overlooked issue: \textit{the mismatch of activation magnitudes at the pruning interface}. The pre-interface activations exhibit significantly different scales from the post-interface ones, causing the distributional shift as it propagates through the remaining layers. To address this issue, we introduce \textsc{LinearPatch}, a lightweight and plug-and-play technique that fuses two operations into one matrix multiply at the pruning interface: (i) a Hadamard transformation that suppresses massive outliers at particular tokens and (ii) a channel-wise scaling that aligns activation statistics. On LLaMA-3-8B, \textsc{LinearPatch} preserves up to \textbf{94.15\%} of the original model's performance when pruning 5 out of 32 layers, outperforming the previous state of the art by \textbf{4\%}. The patch can be further refined with 5K unlabeled samples via memory-efficient offline distillation, pushing the retention to 95.16\% within only 30 minutes on a single GPU. Code is available at https://github.com/chenxinrui-tsinghua/LinearPatch.

Xiaohao Liu, Xiaobo Xia, Weixiang Zhao et al.

Large language models (LLMs) have achieved notable progress. Despite their success, next-token prediction (NTP), the dominant method for LLM training and inference, is constrained in both contextual coverage and inference efficiency due to its inherently sequential process. To overcome these challenges, we propose leap multi-token prediction~(L-MTP), an innovative token prediction method that extends the capabilities of multi-token prediction (MTP) by introducing a leap-based mechanism. Unlike conventional MTP, which generates multiple tokens at adjacent positions, L-MTP strategically skips over intermediate tokens, predicting non-sequential ones in a single forward pass. This structured leap not only enhances the model's ability to capture long-range dependencies but also enables a decoding strategy specially optimized for non-sequential leap token generation, effectively accelerating inference. We theoretically demonstrate the benefit of L-MTP in improving inference efficiency. Experiments across diverse benchmarks validate its merit in boosting both LLM performance and inference speed. The source code is available at https://github.com/Xiaohao-Liu/L-MTP.

Yehui Tang, Xiaosong Li, Fangcheng Liu et al.

The surgence of Mixture of Experts (MoE) in Large Language Models promises a small price of execution cost for a much larger model parameter count and learning capacity, because only a small fraction of parameters are activated for each input token. However, it is commonly observed that some experts are activated far more often than others, leading to system inefficiency when running the experts on different devices in parallel. Therefore, we introduce Mixture of Grouped Experts (MoGE), which groups the experts during selection and balances the expert workload better than MoE in nature. It constrains tokens to activate an equal number of experts within each predefined expert group. When a model execution is distributed on multiple devices, this architectural design ensures a balanced computational load across devices, significantly enhancing throughput, particularly for the inference phase. Further, we build Pangu Pro MoE on Ascend NPUs, a sparse model based on MoGE with 72 billion total parameters, 16 billion of which are activated for each token. The configuration of Pangu Pro MoE is optimized for Ascend 300I Duo and 800I A2 through extensive system simulation studies. Our experiments indicate that MoGE indeed leads to better expert load balancing and more efficient execution for both model training and inference on Ascend NPUs. The inference performance of Pangu Pro MoE achieves 1148 tokens/s per card and can be further improved to 1528 tokens/s per card by speculative acceleration, outperforming comparable 32B and 72B Dense models. Furthermore, we achieve an excellent cost-to-performance ratio for model inference on Ascend 300I Duo. Our studies show that Ascend NPUs are capable of training Pangu Pro MoE with massive parallelization to make it a leading model within the sub-100B total parameter class, outperforming prominent open-source models like GLM-Z1-32B and Qwen3-32B.

Anxin Tian, Yiming Li, Xing Li et al.

Agentic memory systems have become critical for enabling LLM agents to maintain long-term context and retrieve relevant information efficiently. However, existing memory frameworks suffer from a fundamental limitation: they perform exhaustive retrieval across the entire storage layer regardless of query characteristics. This brute-force approach creates severe latency bottlenecks as memory grows, hindering real-time agent interactions. We propose SwiftMem, a query-aware agentic memory system that achieves sub-linear retrieval through specialized indexing over temporal and semantic dimensions. Our temporal index enables logarithmic-time range queries for time-sensitive retrieval, while the semantic DAG-Tag index maps queries to relevant topics through hierarchical tag structures. To address memory fragmentation during growth, we introduce an embedding-tag co-consolidation mechanism that reorganizes storage based on semantic clusters to improve cache locality. Experiments on LoCoMo and LongMemEval benchmarks demonstrate that SwiftMem achieves 47$\times$ faster search compared to state-of-the-art baselines while maintaining competitive accuracy, enabling practical deployment of memory-augmented LLM agents.

Qihui Zhu, Tao Zhang, Yuchen Wang et al.

In multimodal large language models (MLLMs), the surge of visual tokens significantly increases the inference time and computational overhead, making them impractical for real-time or resource-constrained applications. Visual token pruning is a promising strategy for reducing the cost of MLLM inference by removing redundant visual tokens. Existing research usually assumes that all attention heads contribute equally to the visual interpretation. However, our study reveals that different heads may capture distinct visual semantics and inherently play distinct roles in visual processing. In light of this observation, we propose HAWK, a head importance-aware visual token pruning method that perceives the varying importance of attention heads in visual tasks to maximize the retention of crucial tokens. By leveraging head importance weights and text-guided attention to assess visual token significance, HAWK effectively retains task-relevant visual tokens while removing redundant ones. The proposed HAWK is entirely training-free and can be seamlessly applied to various MLLMs. Extensive experiments on multiple mainstream vision-language benchmarks demonstrate that HAWK achieves state-of-the-art accuracy. When applied to Qwen2.5-VL, HAWK retains 96.0% of the original accuracy after pruning 80.2% of the visual tokens. Additionally, it reduces end-to-end latency to 74.4% of the original and further decreases GPU memory usage across the tested models. The code is available at https://github.com/peppery77/HAWK.git.

Zehua Pei, Hui-Ling Zhen, Ying Zhang et al.

Current benchmarks for Large Language Models (LLMs) primarily focus on performance metrics, often failing to capture the nuanced behavioral characteristics that differentiate them. This paper introduces a novel ``Behavioral Fingerprinting'' framework designed to move beyond traditional evaluation by creating a multi-faceted profile of a model's intrinsic cognitive and interactive styles. Using a curated \textit{Diagnostic Prompt Suite} and an innovative, automated evaluation pipeline where a powerful LLM acts as an impartial judge, we analyze eighteen models across capability tiers. Our results reveal a critical divergence in the LLM landscape: while core capabilities like abstract and causal reasoning are converging among top models, alignment-related behaviors such as sycophancy and semantic robustness vary dramatically. We further document a cross-model default persona clustering (ISTJ/ESTJ) that likely reflects common alignment incentives. Taken together, this suggests that a model's interactive nature is not an emergent property of its scale or reasoning power, but a direct consequence of specific, and highly variable, developer alignment strategies. Our framework provides a reproducible and scalable methodology for uncovering these deep behavioral differences. Project: https://github.com/JarvisPei/Behavioral-Fingerprinting

Zehua Pei, Lancheng Zou, Hui-Ling Zhen et al.

Scaling large language models (LLMs) improves performance but dramatically increases inference costs. The feed-forward network (FFN), consuming approximately 70\% of inference compute, represents a critical bottleneck, particularly in large batch size scenarios. While mixture-of-experts (MoE) architectures leverage activation sparsity for efficiency, converting existing dense models to MoEs traditionally requires resource-intensive continual pre-training. We present CMoE, a framework that rapidly transforms dense LLMs into MoEs without training. The key innovation lies in analyzing FFN neuron activations to partition them into shared (always active) and routed experts. Routed neurons are clustered using a balanced assignment algorithm, and a differentiable router is constructed analytically from activation statistics, enabling immediate deployment or optional lightweight fine-tuning. Experiments demonstrate that, with activation ratio of 75\%, it achieves remarkable results, delivering lossless precision in terms of perplexity while still maintaining a 5\% acceleration. Further experiments reveal that a CMoE configuration activating just 25\% of parameters reduces end-to-end latency by 1.5x while preserving usable perplexity without additional training. Moreover, a brief LoRA fine-tuning process (requiring only 1 hour and 2,000 samples) successfully recovers over 76\% of the dense model's downstream accuracy. By effectively balancing performance and efficiency, CMoE offers a viable path forward for deploying LLMs in real-world scenarios where computational resources are limited. We make our code publicly available at https://github.com/JarvisPei/CMoE.

Shengyin Sun, Yiming Li, Renxi Liu et al.

Judge Decoding accelerates LLM inference by relaxing the strict verification of Speculative Decoding, yet it typically relies on expensive and noisy supervision. In this work, we revisit this paradigm from first principles, revealing that the ``criticality'' scores learned via costly supervision are intrinsically encoded in the draft-target distributional divergence. We theoretically prove a structural correspondence between learned linear judges and Kullback-Leibler (KL) divergence, demonstrating they rely on the same underlying logit primitives. Guided by this, we propose a simple, training-free verification mechanism based on KL divergence. Extensive experiments across reasoning and coding benchmarks show that our method matches or outperforms complex trained judges (e.g., AutoJudge), offering superior robustness to domain shifts and eliminating the supervision bottleneck entirely.

Xing Li, Hui-Ling Zhen, Lihao Yin et al.

This paper presents a comprehensive empirical study on the safety alignment capabilities. We evaluate what matters for safety alignment in LLMs and LRMs to provide essential insights for developing more secure and reliable AI systems. We systematically investigate and compare the influence of six critical intrinsic model characteristics and three external attack techniques. Our large-scale evaluation is conducted using 32 recent, popular LLMs and LRMs across thirteen distinct model families, spanning a parameter scale from 3B to 235B. The assessment leverages five established safety datasets and probes model vulnerabilities with 56 jailbreak techniques and four CoT attack strategies, resulting in 4.6M API calls. Our key empirical findings are fourfold. First, we identify the LRMs GPT-OSS-20B, Qwen3-Next-80B-A3B-Thinking, and GPT-OSS-120B as the top-three safest models, which substantiates the significant advantage of integrated reasoning and self-reflection mechanisms for robust safety alignment. Second, post-training and knowledge distillation may lead to a systematic degradation of safety alignment. We thus argue that safety must be treated as an explicit constraint or a core optimization objective during these stages, not merely subordinated to the pursuit of general capability. Third, we reveal a pronounced vulnerability: employing a CoT attack via a response prefix can elevate the attack success rate by 3.34x on average and from 0.6% to 96.3% for Seed-OSS-36B-Instruct. This critical finding underscores the safety risks inherent in text-completion interfaces and features that allow user-defined response prefixes in LLM services, highlighting an urgent need for architectural and deployment safeguards. Fourth, roleplay, prompt injection, and gradient-based search for adversarial prompts are the predominant methodologies for eliciting unaligned behaviors in modern models.

Zhongqian Fu, Ning Ding, Kai Han et al.

Mixture-of-Experts (MoE) models have emerged as a cornerstone of large-scale deep learning by efficiently distributing computation and enhancing performance. However, their unique architecture-characterized by sparse expert activation and dynamic routing mechanisms-introduces inherent complexities that challenge conventional quantization techniques. Existing post-training quantization (PTQ) methods struggle to address activation outliers, router consistency and sparse expert calibration, leading to significant performance degradation. To bridge this gap, we propose EAQuant, a novel PTQ framework tailored for MoE architectures. Our method systematically tackles these challenges through three key innovations: (1) expert-aware smoothing aggregation to suppress activation outliers and stabilize quantization, (2) router logits distribution alignment to preserve expert selection consistency post-quantization, and (3) expert-level calibration data balance to optimize sparsely activated experts. Extensive experiments across W4A4 and extreme W3A4 quantization configurations demonstrate that EAQuant significantly outperforms existing methods, achieving average score improvements of 1.15 - 2.28% across three diverse MoE architectures, with particularly pronounced gains in reasoning tasks and robust performance retention under aggressive quantization. By integrating these innovations, EAQuant establishes a new state-of-the-art for high-precision, efficient MoE model compression. Our code is available at https://github.com/darren-fzq1/EAQuant.

Zehua Pei, Ying Zhang, Hui-Ling Zhen et al.

Mixture-of-experts (MoE) architectures enable scaling large language models (LLMs) to vast parameter counts without a proportional rise in computational costs. However, the significant memory demands of large MoE models hinder their deployment across various computational environments, from cloud servers to consumer devices. This study first demonstrates pronounced task-specific specialization in expert activation patterns within MoE layers. Building on this, we introduce PreMoe, a novel framework that enables efficient deployment of massive MoE models in memory-constrained environments. PreMoe features two main components: probabilistic expert pruning (PEP) and task-adaptive expert retrieval (TAER). PEP employs a new metric, the task-conditioned expected selection score (TCESS), derived from router logits to quantify expert importance for specific tasks, thereby identifying a minimal set of critical experts. TAER leverages these task-specific expert importance profiles for efficient inference. It pre-computes and stores compact expert patterns for diverse tasks. When a user query is received, TAER rapidly identifies the most relevant stored task pattern and reconstructs the model by loading only the small subset of experts crucial for that task. This approach dramatically reduces the memory footprint across all deployment scenarios. DeepSeek-R1 671B maintains 97.2\% accuracy on MATH500 when pruned to 8/128 configuration (50\% expert reduction), and still achieves 72.0\% with aggressive 8/32 pruning (87.5\% expert reduction). Pangu-Ultra-MoE 718B achieves 97.15\% on MATH500 and 81.3\% on AIME24 with 8/128 pruning, while even more aggressive pruning to 4/64 (390GB memory) preserves 96.95\% accuracy on MATH500. We make our code publicly available at https://github.com/JarvisPei/PreMoe.

Qingyue Yang, Jie Wang, Xing Li et al.

With the development of large language models (LLMs), efficient inference through Key-Value (KV) cache compression has attracted considerable attention, especially for long-context generation. To compress the KV cache, recent methods identify critical KV tokens through static modeling of attention scores. However, these methods often struggle to accurately determine critical tokens as they neglect the temporal patterns in attention scores, resulting in a noticeable degradation in LLM performance. To address this challenge, we propose AttentionPredictor, which is the first learning-based method to directly predict attention patterns for KV cache compression and critical token identification. Specifically, AttentionPredictor learns a lightweight, unified convolution model to dynamically capture spatiotemporal patterns and predict the next-token attention scores. An appealing feature of AttentionPredictor is that it accurately predicts the attention score and shares the unified prediction model, which consumes negligible memory, among all transformer layers. Moreover, we propose a cross-token critical cache prefetching framework that hides the token estimation time overhead to accelerate the decoding stage. By retaining most of the attention information, AttentionPredictor achieves 13$\times$ KV cache compression and 5.6$\times$ speedup in a cache offloading scenario with comparable LLM performance, significantly outperforming the state-of-the-arts. The code is available at https://github.com/MIRALab-USTC/LLM-AttentionPredictor.

Weizhe Lin, Xing Li, Zhiyuan Yang et al.

Large Reasoning Models (LRMs) demonstrate exceptional capability in tackling complex mathematical, logical, and coding tasks by leveraging extended Chain-of-Thought (CoT) reasoning. Test-time scaling methods, such as prolonging CoT with explicit token-level exploration, can push LRMs' accuracy boundaries, but they incur significant decoding overhead. A key inefficiency source is LRMs often generate redundant thinking CoTs, which demonstrate clear structured overthinking and underthinking patterns. Inspired by human cognitive reasoning processes and numerical optimization theories, we propose TrimR, a verifier-based, training-free, efficient framework for dynamic CoT compression to trim reasoning and enhance test-time scaling, explicitly tailored for production-level deployment. Our method employs a lightweight, pretrained, instruction-tuned verifier to detect and truncate redundant intermediate thoughts of LRMs without any LRM or verifier fine-tuning. We present both the core algorithm and asynchronous online system engineered for high-throughput industrial applications. Empirical evaluations on Ascend NPUs and vLLM show that our framework delivers substantial gains in inference efficiency under large-batch workloads. In particular, on the four MATH500, AIME24, AIME25, and GPQA benchmarks, the reasoning runtime of Pangu Pro MoE, Pangu-R-38B, QwQ-32B, and DeepSeek-R1-Distill-Qwen-32B is improved by up to 70% with negligible impact on accuracy.

Zehua Pei, Hui-Ling Zhen, Xianzhi Yu et al.

Generative Pre-trained Transformers (GPTs) have demonstrated remarkable performance across diverse domains, largely due to the extensive scaling of model parameters. Recent works have observed redundancy within transformer blocks and developed compression methods by structured pruning of less important blocks. However, such direct removal often leads to irreversible performance degradation. In this paper, we propose FuseGPT, a novel methodology designed to recycle pruned transformer blocks, thereby recovering the model's performance. Firstly, we introduce a new importance detection metric, Macro Influence (MI), which evaluates the long-term impact of each transformer block by quantifying the information loss incurred upon its removal. Next, we propose group-level layer fusion, which leverages the parameters from layers of less important blocks and integrates them into the corresponding layers of neighboring blocks. This fusion process is not a one-time operation but is refined through iterative parameter updates by lightweight group-level fine-tuning. Specifically, the injected parameters are frozen but are weighted with learnable rank decomposition matrices to reduce the computational overhead during fine-tuning. Our approach not only works well for large language models but also for large multimodal models. Experimental results indicate that, even with modest amounts of data, FuseGPT surpasses previous methods in both perplexity and zero-shot task performance.

Yehui Tang, Yichun Yin, Yaoyuan Wang et al.

Sparse large language models (LLMs) with Mixture of Experts (MoE) and close to a trillion parameters are dominating the realm of most capable language models. However, the massive model scale poses significant challenges for the underlying software and hardware systems. In this paper, we aim to uncover a recipe to harness such scale on Ascend NPUs. The key goals are better usage of the computing resources under the dynamic sparse model structures and materializing the expected performance gain on the actual hardware. To select model configurations suitable for Ascend NPUs without repeatedly running the expensive experiments, we leverage simulation to compare the trade-off of various model hyperparameters. This study led to Pangu Ultra MoE, a sparse LLM with 718 billion parameters, and we conducted experiments on the model to verify the simulation results. On the system side, we dig into Expert Parallelism to optimize the communication between NPU devices to reduce the synchronization overhead. We also optimize the memory efficiency within the devices to further reduce the parameter and activation management overhead. In the end, we achieve an MFU of 30.0% when training Pangu Ultra MoE, with performance comparable to that of DeepSeek R1, on 6K Ascend NPUs, and demonstrate that the Ascend system is capable of harnessing all the training stages of the state-of-the-art language models. Extensive experiments indicate that our recipe can lead to efficient training of large-scale sparse language models with MoE. We also study the behaviors of such models for future reference.

Haoran Lin, Xianzhi Yu, Kang Zhao et al.

FlashAttention series has been widely applied in the inference of large language models (LLMs). However, FlashAttention series only supports the high-level GPU architectures, e.g., Ampere and Hopper. At present, FlashAttention series is not easily transferrable to NPUs and low-resource GPUs. Moreover, FlashAttention series is inefficient for multi- NPUs or GPUs inference scenarios. In this work, we propose FastAttention which pioneers the adaptation of FlashAttention series for NPUs and low-resource GPUs to boost LLM inference efficiency. Specifically, we take Ascend NPUs and Volta-based GPUs as representatives for designing our FastAttention. We migrate FlashAttention series to Ascend NPUs by proposing a novel two-level tiling strategy for runtime speedup, tiling-mask strategy for memory saving and the tiling-AllReduce strategy for reducing communication overhead, respectively. Besides, we adapt FlashAttention for Volta-based GPUs by redesigning the operands layout in shared memory and introducing a simple yet effective CPU-GPU cooperative strategy for efficient memory utilization. On Ascend NPUs, our FastAttention can achieve a 10.7$\times$ speedup compared to the standard attention implementation. Llama-7B within FastAttention reaches up to 5.16$\times$ higher throughput than within the standard attention. On Volta architecture GPUs, FastAttention yields 1.43$\times$ speedup compared to its equivalents in \texttt{xformers}. Pangu-38B within FastAttention brings 1.46$\times$ end-to-end speedup using FasterTransformer. Coupled with the propose CPU-GPU cooperative strategy, FastAttention supports a maximal input length of 256K on 8 V100 GPUs. All the codes will be made available soon.

Shengyin Sun, Yiming Li, Xing Li et al.

Test-time scaling has emerged as a powerful paradigm for enhancing the reasoning capabilities of large language models (LLMs) by allocating additional computational resources during inference. However, this paradigm is inherently inefficient due to the generation of redundant and repetitive reasoning traces, leading to significant computational overhead. Speculative decoding offers a promising avenue for mitigating this inefficiency, yet its efficacy in the structured, repetition-rich context of test-time scaling remains largely unexplored. To bridge this gap, we introduce the first comprehensive benchmark designed to evaluate speculative decoding methods for accelerating LLM test-time scaling. Our benchmark provides consistent experimental protocols across representative test-time scaling paradigms (e.g., Best-of-N sampling and multi-round thinking), enabling a fair comparison of three major categories of speculative decoding: model-based, training-based, and n-gram-based methods. Extensive experiments reveal that simple n-gram-based methods effectively capture repetitive patterns, demonstrating unique potential in accelerating test-time scaling. This phenomenon demonstrates the value of integrating n-gram-based methods with model-based or training-based approaches to balance acceleration for both repetitive and diverse reasoning in test-time scaling. We hope this benchmark spurs further research on speculative decoding for test-time scaling, enabling faster and more practical reasoning in LLMs through better handling of repetitive and diverse reasoning paths.

Tao Yuan, Haoli Bai, Yinfei Pan et al.

With the increasing size of large language models, layer pruning has gained increased attention as a hardware-friendly approach for model compression. However, existing layer pruning methods struggle to simultaneously address key practical deployment challenges, including performance degradation, high training costs, and limited acceleration. To overcome these limitations, we propose \name, a task-\underline{E}ffective, training-\underline{E}conomical and inference-\underline{E}fficient layer pruning framework. \namespace introduces two key innovations: (1) a differentiable mask optimization method using a Gumbel-TopK sampler, enabling efficient and precise pruning mask search; and (2) an entropy-aware adaptive knowledge distillation strategy that enhances task performance. Extensive experiments over diverse model architectures and benchmarks demonstrate the superiority of our method over state-of-the-art approaches. Notably, \namespace achieves 96\% accuracy, a mere 0.8\% drop from the original model (96.8\%) on MATH-500 when pruning 25\% layers of Qwen3-32B, outperforming existing SOTA (95\%), with a 1.33$\times$ inference speedup by consuming merely 0.5B tokens (0.5\% of the post-training data volume).

Zongle Huang, Lei Zhu, Zongyuan Zhan et al.

Large Language Models (LLMs) have achieved remarkable success across many applications, with Mixture of Experts (MoE) models demonstrating great potential. Compared to traditional dense models, MoEs achieve better performance with less computation. Speculative decoding (SD) is a widely used technique to accelerate LLM inference without accuracy loss, but it has been considered efficient only for dense models. In this work, we first demonstrate that, under medium batch sizes, MoE surprisingly benefits more from SD than dense models. Furthermore, as MoE becomes sparser -- the prevailing trend in MoE designs -- the batch size range where SD acceleration is expected to be effective becomes broader. To quantitatively understand tradeoffs involved in SD, we develop a reliable modeling based on theoretical analyses. While current SD research primarily focuses on improving acceptance rates of algorithms, changes in workload and model architecture can still lead to degraded SD acceleration even with high acceptance rates. To address this limitation, we introduce a new metric 'target efficiency' that characterizes these effects, thus helping researchers identify system bottlenecks and understand SD acceleration more comprehensively. For scenarios like private serving, this work unveils a new perspective to speed up MoE inference, where existing solutions struggle. Experiments on different GPUs show up to 2.29x speedup for Qwen2-57B-A14B at medium batch sizes and validate our theoretical predictions.

Zili Wang, Tianyu Zhang, Haoli Bai et al.

Test-Time Scaling (TTS) has proven effective in improving the performance of Large Language Models (LLMs) during inference. However, existing research has overlooked the efficiency of TTS from a latency-sensitive perspective. Through a latency-aware evaluation of representative TTS methods, we demonstrate that a compute-optimal TTS does not always result in the lowest latency in scenarios where latency is critical. To address this gap and achieve latency-optimal TTS, we propose two key approaches by optimizing the concurrency configurations: (1) branch-wise parallelism, which leverages multiple concurrent inference branches, and (2) sequence-wise parallelism, enabled by speculative decoding. By integrating these two approaches and allocating computational resources properly to each, our latency-optimal TTS enables a 32B model to reach 82.3% accuracy on MATH-500 within 1 minute and a smaller 3B model to achieve 72.4% within 10 seconds. Our work emphasizes the importance of latency-aware TTS and demonstrates its ability to deliver both speed and accuracy in latency-sensitive scenarios.

Jiansong Li, Xiao Dong, Guangli Li et al.

The training of deep neural networks (DNNs) is usually memory-hungry due to the limited device memory capacity of DNN accelerators. Characterizing the memory behaviors of DNN training is critical to optimize the device memory pressures. In this work, we pinpoint the memory behaviors of each device memory block of GPU during training by instrumenting the memory allocators of the runtime system. Our results show that the memory access patterns of device memory blocks are stable and follow an iterative fashion. These observations are useful for the future optimization of memory-efficient training from the perspective of raw memory access patterns.