Serving Large Language Models on Huawei CloudMatrix384

The rapid evolution of large language models (LLMs), driven by growing parameter scales, adoption of mixture-of-experts (MoE) architectures, and expanding context lengths, imposes unprecedented demands on AI infrastructure. Traditional AI clusters face limitations in compute intensity, memory bandwidth, inter-chip communication, and latency, compounded by variable workloads and strict service-level objectives. Addressing these issues requires fundamentally redesigned hardware-software integration. This paper introduces Huawei CloudMatrix, a next-generation AI datacenter architecture, realized in the production-grade CloudMatrix384 supernode. It integrates 384 Ascend 910C NPUs and 192 Kunpeng CPUs interconnected via an ultra-high-bandwidth Unified Bus (UB) network, enabling direct all-to-all communication and dynamic pooling of resources. These features optimize performance for communication-intensive operations, such as large-scale MoE expert parallelism and distributed key-value cache access. To fully leverage CloudMatrix384, we propose CloudMatrix-Infer, an advanced LLM serving solution incorporating three core innovations: a peer-to-peer serving architecture that independently scales prefill, decode, and caching; a large-scale expert parallelism strategy supporting EP320 via efficient UB-based token dispatch; and hardware-aware optimizations including specialized operators, microbatch-based pipelining, and INT8 quantization. Evaluation with the DeepSeek-R1 model shows CloudMatrix-Infer achieves state-of-the-art efficiency: prefill throughput of 6,688 tokens/s per NPU and decode throughput of 1,943 tokens/s per NPU (<50 ms TPOT). It effectively balances throughput and latency, sustaining 538 tokens/s even under stringent 15 ms latency constraints, while INT8 quantization maintains model accuracy across benchmarks.

Learning Unified Force and Position Control for Legged Loco-Manipulation

Robotic loco-manipulation tasks often involve contact-rich interactions with the environment, requiring the joint modeling of contact force and robot position. However, recent visuomotor policies often focus solely on learning position or force control, overlooking their co-learning. In this work, we propose the first unified policy for legged robots that jointly models force and position control learned without reliance on force sensors. By simulating diverse combinations of position and force commands alongside external disturbance forces, we use reinforcement learning to learn a policy that estimates forces from historical robot states and compensates for them through position and velocity adjustments. This policy enables a wide range of manipulation behaviors under varying force and position inputs, including position tracking, force application, force tracking, and compliant interactions. Furthermore, we demonstrate that the learned policy enhances trajectory-based imitation learning pipelines by incorporating essential contact information through its force estimation module, achieving approximately 39.5% higher success rates across four challenging contact-rich manipulation tasks compared to position-control policies. Extensive experiments on both a quadrupedal manipulator and a humanoid robot validate the versatility and robustness of the proposed policy across diverse scenarios.

InstruGen: Automatic Instruction Generation for Vision-and-Language Navigation Via Large Multimodal Models

Recent research on Vision-and-Language Navigation (VLN) indicates that agents suffer from poor generalization in unseen environments due to the lack of realistic training environments and high-quality path-instruction pairs. Most existing methods for constructing realistic navigation scenes have high costs, and the extension of instructions mainly relies on predefined templates or rules, lacking adaptability. To alleviate the issue, we propose InstruGen, a VLN path-instruction pairs generation paradigm. Specifically, we use YouTube house tour videos as realistic navigation scenes and leverage the powerful visual understanding and generation abilities of large multimodal models (LMMs) to automatically generate diverse and high-quality VLN path-instruction pairs. Our method generates navigation instructions with different granularities and achieves fine-grained alignment between instructions and visual observations, which was difficult to achieve with previous methods. Additionally, we design a multi-stage verification mechanism to reduce hallucinations and inconsistency of LMMs. Experimental results demonstrate that agents trained with path-instruction pairs generated by InstruGen achieves state-of-the-art performance on the R2R and RxR benchmarks, particularly in unseen environments. Code is available at https://github.com/yanyu0526/InstruGen.

LinBridge: A Learnable Framework for Interpreting Nonlinear Neural Encoding Models

Neural encoding of artificial neural networks (ANNs) links their computational representations to brain responses, offering insights into how the brain processes information. Current studies mostly use linear encoding models for clarity, even though brain responses are often nonlinear. This has sparked interest in developing nonlinear encoding models that are still interpretable. To address this problem, we propose LinBridge, a learnable and flexible framework based on Jacobian analysis for interpreting nonlinear encoding models. LinBridge posits that the nonlinear mapping between ANN representations and neural responses can be factorized into a linear inherent component that approximates the complex nonlinear relationship, and a mapping bias that captures sample-selective nonlinearity. The Jacobian matrix, which reflects output change rates relative to input, enables the analysis of sample-selective mapping in nonlinear models. LinBridge employs a self-supervised learning strategy to extract both the linear inherent component and nonlinear mapping biases from the Jacobian matrices of the test set, allowing it to adapt effectively to various nonlinear encoding models. We validate the LinBridge framework in the scenario of neural visual encoding, using computational visual representations from CLIP-ViT to predict brain activity recorded via functional magnetic resonance imaging (fMRI). Our experimental results demonstrate that: 1) the linear inherent component extracted by LinBridge accurately reflects the complex mappings of nonlinear neural encoding models; 2) the sample-selective mapping bias elucidates the variability of nonlinearity across different levels of the visual processing hierarchy. This study presents a novel tool for interpreting nonlinear neural encoding models and offers fresh evidence about hierarchical nonlinearity distribution in the visual cortex.

Governing equation discovery of a complex system from snapshots

Complex systems in physics, chemistry, and biology that evolve over time with inherent randomness are typically described by stochastic differential equations (SDEs). A fundamental challenge in science and engineering is to determine the governing equations of a complex system from snapshot data. Traditional equation discovery methods often rely on stringent assumptions, such as the availability of the trajectory information or time-series data, and the presumption that the underlying system is deterministic. In this work, we introduce a data-driven, simulation-free framework, called Sparse Identification of Differential Equations from Snapshots (SpIDES), that discovers the governing equations of a complex system from snapshots by utilizing the advanced machine learning techniques to perform three essential steps: probability flow reconstruction, probability density estimation, and Bayesian sparse identification. We validate the effectiveness and robustness of SpIDES by successfully identifying the governing equation of an over-damped Langevin system confined within two potential wells. By extracting interpretable drift and diffusion terms from the SDEs, our framework provides deeper insights into system dynamics, enhances predictive accuracy, and facilitates more effective strategies for managing and simulating stochastic systems.

P/D-Serve: Serving Disaggregated Large Language Model at Scale

Serving disaggregated large language models (LLMs) over tens of thousands of xPU devices (GPUs or NPUs) with reliable performance faces multiple challenges. 1) Ignoring the diversity (various prefixes and tidal requests), treating all the prompts in a mixed pool is inadequate. To facilitate the similarity per scenario and minimize the inner mismatch on P/D (prefill and decoding) processing, fine-grained organization is required, dynamically adjusting P/D ratios for better performance. 2) Due to inaccurate estimation on workload (queue status or maintained connections), the global scheduler easily incurs unnecessary timeouts in prefill. 3) Block-fixed device-to-device (D2D) KVCache transfer over cluster-level RDMA (remote direct memory access) fails to achieve desired D2D utilization as expected. To overcome previous problems, this paper proposes an end-to-end system P/D-Serve, complying with the paradigm of MLOps (machine learning operations), which models end-to-end (E2E) P/D performance and enables: 1) fine-grained P/D organization, mapping the service with RoCE (RDMA over converged ethernet) as needed, to facilitate similar processing and dynamic adjustments on P/D ratios; 2) on-demand forwarding upon rejections for idle prefill, decoupling the scheduler from regular inaccurate reports and local queues, to avoid timeouts in prefill; and 3) efficient KVCache transfer via optimized D2D access. P/D-Serve is implemented upon Ascend and MindSpore, has been deployed over tens of thousands of NPUs for more than eight months in commercial use, and further achieves 60\%, 42\% and 46\% improvements on E2E throughput, time-to-first-token (TTFT) SLO (service level objective) and D2D transfer time. As the E2E system with optimizations, P/D-Serve achieves 6.7x increase on throughput, compared with aggregated LLMs.

Risk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems

Large language models (LLMs) have strong capabilities in solving diverse natural language processing tasks. However, the safety and security issues of LLM systems have become the major obstacle to their widespread application. Many studies have extensively investigated risks in LLM systems and developed the corresponding mitigation strategies. Leading-edge enterprises such as OpenAI, Google, Meta, and Anthropic have also made lots of efforts on responsible LLMs. Therefore, there is a growing need to organize the existing studies and establish comprehensive taxonomies for the community. In this paper, we delve into four essential modules of an LLM system, including an input module for receiving prompts, a language model trained on extensive corpora, a toolchain module for development and deployment, and an output module for exporting LLM-generated content. Based on this, we propose a comprehensive taxonomy, which systematically analyzes potential risks associated with each module of an LLM system and discusses the corresponding mitigation strategies. Furthermore, we review prevalent benchmarks, aiming to facilitate the risk assessment of LLM systems. We hope that this paper can help LLM participants embrace a systematic perspective to build their responsible LLM systems.

Lifting by Image -- Leveraging Image Cues for Accurate 3D Human Pose Estimation

The "lifting from 2D pose" method has been the dominant approach to 3D Human Pose Estimation (3DHPE) due to the powerful visual analysis ability of 2D pose estimators. Widely known, there exists a depth ambiguity problem when estimating solely from 2D pose, where one 2D pose can be mapped to multiple 3D poses. Intuitively, the rich semantic and texture information in images can contribute to a more accurate "lifting" procedure. Yet, existing research encounters two primary challenges. Firstly, the distribution of image data in 3D motion capture datasets is too narrow because of the laboratorial environment, which leads to poor generalization ability of methods trained with image information. Secondly, effective strategies for leveraging image information are lacking. In this paper, we give new insight into the cause of poor generalization problems and the effectiveness of image features. Based on that, we propose an advanced framework. Specifically, the framework consists of two stages. First, we enable the keypoints to query and select the beneficial features from all image patches. To reduce the keypoints attention to inconsequential background features, we design a novel Pose-guided Transformer Layer, which adaptively limits the updates to unimportant image patches. Then, through a designed Adaptive Feature Selection Module, we prune less significant image patches from the feature map. In the second stage, we allow the keypoints to further emphasize the retained critical image features. This progressive learning approach prevents further training on insignificant image features. Experimental results show that our model achieves state-of-the-art performance on both the Human3.6M dataset and the MPI-INF-3DHP dataset.