Core Insights - DeepSeek introduced a new architecture called "Manifold-Constrained Hyper-Connections" (mHC), which enhances performance with only a 6.7% increase in training time on a 27 billion parameter model [3][36]. - The mHC architecture optimizes the residual connection space by projecting matrices onto constrained manifolds, ensuring stability and significantly expanding the residual stream width without substantial computational costs [8][25]. Group 1: Performance Improvements - In system-level benchmark tests, the mHC architecture consistently outperformed baseline models and Hyper-Connections (HC) across various tasks, demonstrating its effectiveness in large-scale pre-training [22][51]. - Specific performance metrics showed that mHC achieved a 2.1% improvement on the BBH benchmark and a 2.3% improvement on the DROP benchmark compared to HC [52][54]. Group 2: Technical Details - The core idea of mHC is to restore identity mapping properties under the topology of Hyper-Connections, allowing for practical value in large-scale training and real-world foundational model tasks [25]. - mHC employs a double stochastic matrix constraint to maintain stability while enhancing the interaction between residual streams, which is crucial for maximizing the potential of multi-stream architectures [26][27]. Group 3: Engineering Optimizations - The implementation of mHC involved several engineering optimizations, including reordering operations to improve efficiency and using mixed precision strategies to maximize numerical accuracy without sacrificing computational speed [38][42]. - The DualPipe scheduling strategy was enhanced to effectively overlap communication and computation, addressing significant communication delays introduced by the n-stream residual structure [46][48].

Core Insights - DeepSeek has introduced a new architecture called Manifold-Constrained Hyper-Connections (mHC) to address the instability issues in traditional hyper-connections during large-scale model training while maintaining significant performance gains [1][6][8]. Group 1: mHC Architecture - The mHC architecture extends the single residual flow of traditional Transformers into a multi-flow parallel structure, utilizing the Sinkhorn-Knopp algorithm to constrain the connection matrix on a doubly stochastic matrix manifold [1][8]. - The core objective of mHC is to retain the performance improvements from widening the residual flow while resolving training instability and excessive memory consumption [8][9]. - Empirical evidence shows that mHC not only addresses stability issues but also demonstrates exceptional scalability in large-scale training, such as with a 27 billion parameter model, where it only increased training time by 6.7% while achieving significant performance improvements [8][32]. Group 2: Challenges with Traditional Hyper-Connections - Traditional hyper-connections (HC) have led to severe training instability and limited scalability due to the fundamental disruption of the inherent identity mapping property, which is crucial for stable training [5][9]. - The widening of information channels in HC results in increased memory access overhead, contributing to what is known as the "memory wall" problem [9][5]. Group 3: Implementation and Efficiency - DeepSeek has designed a tailored infrastructure for mHC, which includes kernel fusion, selective recomputation, and an extended DualPipe communication overlap strategy to minimize memory usage and enhance efficiency [23][25][27]. - The Sinkhorn-Knopp algorithm is employed to ensure that the residual connection matrix remains stable and adheres to the properties of a doubly stochastic matrix, which helps mitigate gradient explosion issues [16][21]. Group 4: Experimental Validation - The research team conducted experiments using language model pre-training to validate the effectiveness of mHC, comparing it against baseline models and traditional HC [28][32]. - Results from various downstream benchmark tests indicate that mHC consistently outperforms baseline models and often surpasses HC, demonstrating its effectiveness in large-scale pre-training [34][33]. - The scalability experiments reveal that mHC maintains performance advantages even at higher computational budgets, showing only slight degradation in performance [36][37].

来源：量子位 | 公众号 QbitAI 残差连接十年未变，扩展之后却带来隐患 2026年新年第一天，DeepSeek上传新论文。 给何恺明2016成名作ResNet中提出的深度学习基础组件"残差连接"来了一场新时代的升级。 DeepSeek梁文峰亲自署名论文，共同一作为Zhenda Xie ， Yixuan Wei， Huanqi Cao。 DeepSeek团队的实验表明，在这三个映射中，负责残差流内部信息交换的Hres矩阵贡献了最显著的性能 提升。 残差连接自2016年ResNet问世以来，一直是深度学习架构的基石。 其核心机制简洁明了，x+1 = x + F （x ，W），即下一层的输出等于当前层输入加上残差函数的输 出。 这个设计之所以成功，关键在于"恒等映射"属性，信号可以从浅层直接传递到深层，不经任何修改。 随着Transformer架构的崛起，这一范式已成为GPT、LLaMA等大语言模型的标准配置。 这个设计之所以成功，关键在于"恒等映射"属性，信号可以从浅层直接传递到深层，不经任何修改。 近期出现的Hyper-Connections（HC）试图打破这一格局。HC将残差流的宽度从C维扩展到n×C维 ...

Core Insights - DeepSeek has introduced a new architecture called Manifold-Constrained Hyper-Connections (mHC) aimed at addressing the instability issues in traditional hyper-connections during large-scale model training while maintaining significant performance gains [1][27][28]. Group 1: Architecture and Methodology - The mHC architecture expands the traditional single residual flow of Transformers into a multi-flow parallel structure, utilizing the Sinkhorn-Knopp algorithm to constrain the connection matrix on a doubly stochastic matrix manifold [1][28]. - The core objective of mHC is to retain the performance improvements from widening the residual flow while resolving issues related to training instability and excessive memory consumption [4][34]. - The research team has implemented infrastructure optimizations such as kernel fusion, selective recomputation, and an extended DualPipe communication strategy to offset the overhead caused by wider channels [31][34]. Group 2: Performance and Stability - Empirical evidence shows that mHC not only resolves stability issues but also demonstrates exceptional scalability in large-scale training scenarios, such as with a 27 billion parameter model, where it only increased training time overhead by 6.7% while achieving significant performance improvements [34][49]. - The training stability of mHC was evaluated against a baseline model, showing a reduction in final loss by 0.021 and maintaining a stable gradient norm profile, indicating superior stability compared to traditional hyper-connections [49][50]. Group 3: Benchmarking and Results - In various downstream benchmark tests, mHC consistently outperformed the baseline model and surpassed traditional hyper-connections in most tasks, achieving performance gains of 2.1% and 2.3% in specific tasks [51][52]. - The scalability experiments indicated that mHC maintains its performance advantages even under higher computational budgets, demonstrating robust effectiveness in large-scale scenarios [52][53].

Core Viewpoint - The article discusses the evolution and enhancement of the residual connection, a fundamental component in deep learning introduced by He Kaiming in ResNet, and presents a new approach called Hyper-Connections (HC) that aims to improve performance while addressing potential issues related to signal amplification and stability in deep learning architectures [2][7][11]. Group 1: Residual Connections and Their Evolution - Residual connections have been a cornerstone of deep learning since the introduction of ResNet in 2016, allowing signals to pass directly from shallow to deep layers without modification [7][9]. - The rise of Transformer architectures has made residual connections a standard feature in large language models like GPT and LLaMA [10]. - Hyper-Connections (HC) expand the residual flow width from C dimensions to n×C dimensions, introducing three learnable mapping matrices to manage information flow [11]. Group 2: Performance and Stability Challenges - Experiments by the DeepSeek team indicate that the Hres matrix, responsible for internal information exchange in HC, significantly enhances performance [12]. - However, when HC is extended to multiple layers, the composite mapping loses its identity property, leading to potential issues such as sudden loss spikes and gradient fluctuations during training [14]. - The peak amplification factor of signals in HC can reach 3000, which poses risks of signal distortion during inter-layer propagation [16]. Group 3: Theoretical Framework and Constraints - The core idea of the DeepSeek paper is to constrain the residual mapping matrix to a specific manifold formed by double stochastic matrices, which ensures three key theoretical properties: norm preservation, combinatorial closure, and geometric interpretation [17][19]. - The Sinkhorn-Knopp algorithm is employed to project any matrix onto this manifold, effectively reducing the signal amplification issue observed in HC [21]. Group 4: Engineering Optimizations - The paper details the memory access costs associated with expanding the residual flow width, highlighting significant increases in read and write operations for HC compared to standard residual connections [24]. - To mitigate these costs, the team developed infrastructure optimizations, including the TileLang framework for merging operations and specialized kernels for the Sinkhorn-Knopp algorithm [25][26]. - The paper also discusses pipeline parallelism enhancements to overlap computation and communication, improving overall efficiency [27]. Group 5: Experimental Validation - The paper validates the proposed methods on MoE models of sizes 3B, 9B, and 27B, with an expansion rate of n set to 4 [30]. - In the 27B MoE model, the modified HC (mHC) demonstrated a stable training curve, achieving a loss reduction of 0.021 compared to the baseline while maintaining gradient stability [31]. - Performance improvements were noted in downstream tasks, with mHC outperforming both the baseline and HC in various benchmarks [32][35].

Core Viewpoint - The article discusses the introduction of Multiway Dynamic Dense (MUDD) connections as an effective alternative to residual connections in Transformers, significantly enhancing cross-layer information transfer efficiency in deep learning models [1][4]. Background - Residual connections, introduced by Kaiming He in ResNet, have become foundational in deep learning and Transformer LLMs, but they still face limitations in efficient information transfer across layers [1][7]. - MUDD connections dynamically establish cross-layer connections based on the current hidden state, addressing issues like representation collapse and information overload in residual streams [7][8]. Model Architecture - MUDDFormer architecture allows for independent dynamic connections for different information streams (Q, K, V, R), enhancing the model's ability to gather relevant information from previous layers [10][13]. - The introduction of dynamic connections enables the model to adaptively determine the weight of information extracted from previous layers based on the context of each token [11][13]. Experimental Evaluation - MUDDPythia, a model with 2.8 billion parameters, shows performance comparable to larger models (6.9 billion and 12 billion parameters) with only a 0.23% increase in parameters and a 0.4% increase in computation [4][18]. - The MUDDFormer outperforms baseline models like Transformer++ across various model sizes, demonstrating significant computational efficiency improvements [15][17]. Downstream Task Assessment - In downstream tasks, MUDDPythia exhibits higher accuracy in 0-shot and 5-shot evaluations compared to equivalent Pythia models, indicating enhanced contextual learning capabilities [18][20]. - The model achieves a 2.4 times efficiency leap over the 6.9 billion Pythia model and a 4.2 times efficiency leap over the 12 billion Pythia model in specific evaluations [18][20]. Conclusion - MUDDFormer improves residual connections by establishing independent dynamic cross-layer connections for different information streams, enhancing cross-layer interaction and contextual learning capabilities in Transformers [25].