[原文]DeepSeek-AI research@deepseek.com Abstract We present DeepSeek-V3, a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token. To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were thoroughly validated in DeepSeek-V2. Furthermore, DeepSeek-V3 pioneers an auxiliary-loss-free strategy for load balancing and sets a multi-token prediction training objective for stronger performance. We pre-train DeepSeek-V3 on 14.8 trillion diverse and high-quality...
[原文]n-source model capabilities, we scale up our models and introduce DeepSeek-V3, a large Mixture-of-Experts (MoE) model with 671B parameters, of which 37B are activated for each token. With a forward-looking perspective, we consistently strive for strong model performance and economical costs. Therefore, in terms of architecture, DeepSeek-V3 still adopts Multi-head Latent Attention (MLA) (DeepSeek-AI, 2024c ) for efficient inference and DeepSeekMoE (Dai et al., 2024 ) for cost-effective training. These two architectures have been validated in DeepSeek-V2 (DeepSeek-AI, 2024c ) , demonstrating the...
[原文]DualPipe algorithm for efficient pipeline parallelism, which has fewer pipeline bubbles and hides most of the communication during training through computation-communication overlap. This overlap ensures that, as the model further scales up, as long as we maintain a constant computation-to-communication ratio, we can still employ fine-grained experts across nodes while achieving a near-zero all-to-all communication overhead. In addition, we also develop efficient cross-node all-to-all communication kernels to fully utilize InfiniBand (IB) and NVLink bandwidths. Furthermore, we meticulously opt...
[原文]rce models, including GPT-4o and Claude-3.5-Sonnet, on a series of standard and open-ended benchmarks. Training Costs Pre-Training Context Extension Post-Training Total in H800 GPU Hours 2664K 119K 5K 2788K in USD $5.328M $0.238M $0.01M $5.576M Table 1: Training costs of DeepSeek-V3, assuming the rental price of H800 is $2 per GPU hour. Lastly, we emphasize again the economical training costs of DeepSeek-V3, summarized in Table 1 , achieved through our optimized co-design of algorithms, frameworks, and hardware. During the pre-training stage, training DeepSeek-V3 on each trillion tokens requir...
[原文]hms, frameworks, and hardware, we overcome the communication bottleneck in cross-node MoE training, achieving near-full computation-communication overlap. This significantly enhances our training efficiency and reduces the training costs, enabling us to further scale up the model size without additional overhead. • At an economical cost of only 2.664M H800 GPU hours, we complete the pre-training of DeepSeek-V3 on 14.8T tokens, producing the currently strongest open-source base model. The subsequent training stages after pre-training require only 0.1M GPU hours. Post-Training: Knowledge Distill...
[原文]d-source models. Notably, it even outperforms o1-preview on specific benchmarks, such as MATH-500, demonstrating its robust mathematical reasoning capabilities. (2) On coding-related tasks, DeepSeek-V3 emerges as the top-performing model for coding competition benchmarks, such as LiveCodeBench, solidifying its position as the leading model in this domain. For engineering-related tasks, while DeepSeek-V3 performs slightly below Claude-Sonnet-3.5, it still outpaces all other models by a significant margin, demonstrating its competitiveness across diverse technical benchmarks. In the remainder of...
[原文]f the basic architecture of DeepSeek-V3. Following DeepSeek-V2, we adopt MLA and DeepSeekMoE for efficient inference and economical training. 2.1 Basic Architecture The basic architecture of DeepSeek-V3 is still within the Transformer (Vaswani et al., 2017 ) framework. For efficient inference and economical training, DeepSeek-V3 also adopts MLA and DeepSeekMoE, which have been thoroughly validated by DeepSeek-V2. Compared with DeepSeek-V2, an exception is that we additionally introduce an auxiliary-loss-free load balancing strategy (Wang et al., 2024a ) for DeepSeekMoE to mitigate the performa...
[原文]In recent years, Large Language Models (LLMs) have been undergoing rapid iteration and evolution (OpenAI, 2024a ; Anthropic, 2024 ; Google, 2024 ) , progressively diminishing the gap towards Artificial General Intelligence (AGI). Beyond closed-source models, open-source models, including DeepSeek series (DeepSeek-AI, 2024b , c ; Guo et al., 2024 ; DeepSeek-AI, 2024a ) , LLaMA series (Touvron et al., 2023a , b ; AI@Meta, 2024a , b ) , Qwen series (Qwen, 2023 , 2024a , 2024b ) , and Mistral series (Jiang et al., 2023 ; Mistral, 2024 ) , are also making significant strides, endeavoring to close t...
1 Introduction
训练框架的优化。低精度训练已成为实现高效训练的一种极具前景的解决方案（Kalamkar et al., 2019 ; Narang et al., 2017 ; Peng et al., 2023b ; Dettmers et al., 2022），其发展与硬件能力的提升密切相关（Micikevicius et al., 2022 ; Luo et al., 2024 ; Rouhani et al., 2023a）。在本工作中，我们提出了一种 FP8 混合精度训练框架，并首次在超大规模模型上验证了其有效性。通过对 FP8 计算与存储的支持，我们同时实现了加速的……
[原文]mizations for the training framework. Low-precision training has emerged as a promising solution for efficient training (Kalamkar et al., 2019 ; Narang et al., 2017 ; Peng et al., 2023b ; Dettmers et al., 2022 ) , its evolution being closely tied to advancements in hardware capabilities (Micikevicius et al., 2022 ; Luo et al., 2024 ; Rouhani et al., 2023a ) . In this work, we introduce an FP8 mixed precision training framework and, for the first time, validate its effectiveness on an extremely large-scale model. Through the support for FP8 computation and storage, we achieve both accelerated t...
[原文](RL) on the base model of DeepSeek-V3, to align it with human preferences and further unlock its potential. During the post-training stage, we distill the reasoning capability from the DeepSeek-R1 series of models, and meanwhile carefully maintain the balance between model accuracy and generation length. We evaluate DeepSeek-V3 on a comprehensive array of benchmarks. Despite its economical training costs, comprehensive evaluations reveal that DeepSeek-V3-Base has emerged as the strongest open-source base model currently available, especially in code and math. Its chat version also outperforms ...
[原文]alancing Strategy and Training Objective • On top of the efficient architecture of DeepSeek-V2, we pioneer an auxiliary-loss-free strategy for load balancing, which minimizes the performance degradation that arises from encouraging load balancing. • We investigate a Multi-Token Prediction (MTP) objective and prove it beneficial to model performance. It can also be used for speculative decoding for inference acceleration. Pre-Training: Towards Ultimate Training Efficiency • We design an FP8 mixed precision training framework and, for the first time, validate the feasibility and effectiveness of...
[原文]ce is comparable to leading closed-source models like GPT-4o and Claude-Sonnet-3.5, narrowing the gap between open-source and closed-source models in this domain. (2) For factuality benchmarks, DeepSeek-V3 demonstrates superior performance among open-source models on both SimpleQA and Chinese SimpleQA. While it trails behind GPT-4o and Claude-Sonnet-3.5 in English factual knowledge (SimpleQA), it surpasses these models in Chinese factual knowledge (Chinese SimpleQA), highlighting its strength in Chinese factual knowledge. • Code, Math, and Reasoning : (1) DeepSeek-V3 achieves state-of-the-art ...
[原文]Section 5 ). Lastly, we conclude this work, discuss existing limitations of DeepSeek-V3, and propose potential directions for future research (Section 6 ).
[原文]We first introduce the basic architecture of DeepSeek-V3, featured by Multi-head Latent Attention (MLA) (DeepSeek-AI, 2024c ) for efficient inference and DeepSeekMoE (Dai et al., 2024 ) for economical training. Then, we present a Multi-Token Prediction (MTP) training objective, which we have observed to enhance the overall performance on evaluation benchmarks. For other minor details not explicitly mentioned, DeepSeek-V3 adheres to the settings of DeepSeek-V2 (DeepSeek-AI, 2024c ) . Figure 2: Illustration of the basic architecture of DeepSeek-V3. Following DeepSeek-V2, we adopt MLA and DeepSee...
[原文]h C ] = 𝐤 t C \displaystyle[\mathbf{k}_{t,1}^{C};\mathbf{k}_{t,2}^{C};...;\mathbf{k}_{t,n_{h}}^{C}]=\mathbf{k}_{t}^{C} = W U K 𝐜 t K V , \displaystyle=W^{UK}\mathbf{c}_{t}^{KV}, (2) 𝐤 t R \displaystyle\boxed{\color[rgb]{0,0,1}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,1}\mathbf{k}_{t}^{R}} = RoPE ( W K R 𝐡 t ) , \displaystyle=\operatorname{RoPE}({W^{KR}}\mathbf{h}_{t}), (3) 𝐤 t , i \displaystyle\mathbf{k}_{t,i} = [ 𝐤 t , i C ; 𝐤 t R ] , \displaystyle=[\mathbf{k}_{t,i}^{C};\mathbf{k}_{t}^{R}], (4) [ 𝐯 t , 1 C ; 𝐯 t , 2 C ; … ; 𝐯 t , n h C ] = 𝐯 t C \displaystyle[\mathbf{v}_{t,1}^{...
[原文]o}_{t,n_{h}}], (11) where W O ∈ ℝ d × d h n h W^{O}\in\mathbb{R}^{d\times d_{h}n_{h}} denotes the output projection matrix. 2.1.2 DeepSeekMoE with Auxiliary-Loss-Free Load Balancing Basic Architecture of DeepSeekMoE. For Feed-Forward Networks (FFNs), DeepSeek-V3 employs the DeepSeekMoE architecture (Dai et al., 2024 ) . Compared with traditional MoE architectures like GShard (Lepikhin et al., 2021 ) , DeepSeekMoE uses finer-grained experts and isolates some experts as shared ones. Let 𝐮 t \mathbf{u}_{t} denote the FFN input of the t t -th token, we compute the FFN output 𝐡 t ′ \mathbf{h}_{t}...
[原文]_{i} is the centroid vector of the i i -th routed expert; and Topk ( ⋅ , K ) \operatorname{Topk}(\cdot,K) denotes the set comprising K K highest scores among the affinity scores calculated for the t t -th token and all routed experts. Slightly different from DeepSeek-V2, DeepSeek-V3 uses the sigmoid function to compute the affinity scores, and applies a normalization among all selected affinity scores to produce the gating values. Auxiliary-Loss-Free Load Balancing. For MoE models, an unbalanced expert load will lead to routing collapse (Shazeer et al., 2017 ) and diminish computational effi...
[原文]arameter called bias update speed. Through the dynamic adjustment, DeepSeek-V3 keeps balanced expert load during training, and achieves better performance than models that encourage load balance through pure auxiliary losses. Complementary Sequence-Wise Auxiliary Loss. Although DeepSeek-V3 mainly relies on the auxiliary-loss-free strategy for load balance, to prevent extreme imbalance within any single sequence, we also employ a complementary sequence-wise balance loss: ℒ Bal \displaystyle\mathcal{L}_{\mathrm{Bal}} = α ∑ i = 1 N r f i P i , \displaystyle=\alpha\sum_{i=1}^{N_{r}}{f_{i}P_{i}...
[原文]y, DeepSeek-V3 keeps a good load balance during its full training. Therefore, DeepSeek-V3 does not drop any tokens during training. In addition, we also implement specific deployment strategies to ensure inference load balance, so DeepSeek-V3 also does not drop tokens during inference. Figure 3: Illustration of our Multi-Token Prediction (MTP) implementation. We keep the complete causal chain for the prediction of each token at each depth. 2.2 Multi-Token Prediction Inspired by Gloeckle et al. ( 2024 ) , we investigate and set a Multi-Token Prediction (MTP) objective for DeepSeek-V3, which ext...
MTP 模块的训练损失定义为：\mathcal{L}_{\text{MTP}}^{k}=\operatorname{CrossEntropy}(P_{2+k:T+1}^{k},t_{2+k:T+1})=-\frac{1}{T}\sum_{i=2+k}^{T+1}\log P_{i}^{k}[t_{i}]，(24) 其中 T 表示输入序列长度，t_{i} 表示第 i 个位置的真实 token，P_{i}^{k}[t_{i}] 表示第 k 个 MTP 模块给出的对应预测概率。最后，我们计算所有深度的 MTP 损失平均值，并乘以权重系数 \lambda 加入总训练损失。多 Token 预测模块通过同时预测多个未来 token，增强了训练信号密度，并提升了推理时的 speculative decoding 效率。
[原文]: ℒ MTP k = CrossEntropy ( P 2 + k : T + 1 k , t 2 + k : T + 1 ) = − 1 T ∑ i = 2 + k T + 1 log P i k [ t i ] , \mathcal{L}_{\text{MTP}}^{k}=\operatorname{CrossEntropy}(P_{2+k:T+1}^{k},t_{2+k:T+1})=-\frac{1}{T}\sum_{i=2+k}^{T+1}\log P_{i}^{k}[t_{i}], (24) where T T denotes the input sequence length, t i t_{i} denotes the ground-truth token at the i i -th position, and P i k [ t i ] P_{i}^{k}[t_{i}] denotes the corresponding prediction probability of t i t_{i} , given by the k k -th MTP module. Finally, we compute the average of the MTP losses across all depths and multiply it by a wei...
[原文]The basic architecture of DeepSeek-V3 is still within the Transformer (Vaswani et al., 2017 ) framework. For efficient inference and economical training, DeepSeek-V3 also adopts MLA and DeepSeekMoE, which have been thoroughly validated by DeepSeek-V2. Compared with DeepSeek-V2, an exception is that we additionally introduce an auxiliary-loss-free load balancing strategy (Wang et al., 2024a ) for DeepSeekMoE to mitigate the performance degradation induced by the effort to ensure load balance. Figure 2 illustrates the basic architecture of DeepSeek-V3, and we will briefly review the details of M...
[原文]U V 𝐜 t K V , \displaystyle=W^{UV}\mathbf{c}_{t}^{KV}, (5) where 𝐜 t K V ∈ ℝ d c \mathbf{c}_{t}^{KV}\in\mathbb{R}^{d_{c}} is the compressed latent vector for keys and values; d c ( ≪ d h n h ) d_{c}(\ll d_{h}n_{h}) indicates the KV compression dimension; W D K V ∈ ℝ d c × d W^{DKV}\in\mathbb{R}^{d_{c}\times d} denotes the down-projection matrix; W U K , W U V ∈ ℝ d h n h × d c W^{UK},W^{UV}\in\mathbb{R}^{d_{h}n_{h}\times d_{c}} are the up-projection matrices for keys and values, respectively; W K R ∈ ℝ d h R × d W^{KR}\in\mathbb{R}^{d_{h}^{R}\times d} is the matrix used t...
[原文]}_{t,i} = [ 𝐪 t , i C ; 𝐪 t , i R ] , \displaystyle=[\mathbf{q}_{t,i}^{C};\mathbf{q}_{t,i}^{R}], (9) where 𝐜 t Q ∈ ℝ d c ′ \mathbf{c}_{t}^{Q}\in\mathbb{R}^{d_{c}^{\prime}} is the compressed latent vector for queries; d c ′ ( ≪ d h n h ) d_{c}^{\prime}(\ll d_{h}n_{h}) denotes the query compression dimension; W D Q ∈ ℝ d c ′ × d , W U Q ∈ ℝ d h n h × d c ′ W^{DQ}\in\mathbb{R}^{d_{c}^{\prime}\times d},W^{UQ}\in\mathbb{R}^{d_{h}n_{h}\times d_{c}^{\prime}} are the down-projection and up-projection matrices for queries, respectively; and W Q R ∈ ℝ d h R n h × d c ′ W^{QR}\in\mathbb{R}^{d...
[原文]ely on the auxiliary loss (Fedus et al., 2021 ; Lepikhin et al., 2021 ) to avoid unbalanced load. However, too large an auxiliary loss will impair the model performance (Wang et al., 2024a ) . To achieve a better trade-off between load balance and model performance, we pioneer an auxiliary-loss-free load balancing strategy (Wang et al., 2024a ) to ensure load balance. To be specific, we introduce a bias term b i b_{i} for each expert and add it to the corresponding affinity scores s i , t s_{i,t} to determine the top-K routing: g i , t ′ \displaystyle g^{\prime}_{i,t} = { s i , t , s i , t + b...
[原文]}T}\sum_{t=1}^{T}\mathds{1} ( s i , t ∈ Topk ( { s j , t | 1 ⩽ j ⩽ N r } , K r ) ) , \displaystyle\left(s_{i,t}\in\operatorname{Topk}(\{s_{j,t}|1\leqslant j\leqslant N_{r}\},K_{r})\right), (18) s i , t ′ \displaystyle s^{\prime}_{i,t} = s i , t ∑ j = 1 N r s j , t , \displaystyle=\frac{s_{i,t}}{\sum_{j=1}^{N_{r}}s_{j,t}}, (19) P i \displaystyle P_{i} = 1 T ∑ t = 1 T s i , t ′ , \displaystyle=\frac{1}{T}\sum_{t=1}^{T}{s^{\prime}_{i,t}}, (20) where the balance factor α \alpha is a hyper-parameter, which will be assigned an extremely small value for DeepSeek-V3; 𝟙 ( ⋅ ) \mathds{1}(\cdot) de...
[原文]Inspired by Gloeckle et al. ( 2024 ) , we investigate and set a Multi-Token Prediction (MTP) objective for DeepSeek-V3, which extends the prediction scope to multiple future tokens at each position. On the one hand, an MTP objective densifies the training signals and may improve data efficiency. On the other hand, MTP may enable the model to pre-plan its representations for better prediction of future tokens. Figure 3 illustrates our implementation of MTP. Different from Gloeckle et al. ( 2024 ) , which parallelly predicts D D additional tokens using independent output heads, we sequentially p...
[原文]. The combined 𝐡 i ′ k \mathbf{h}_{i}^{\prime k} serves as the input of the Transformer block at the k k -th depth to produce the output representation at the current depth 𝐡 i k \mathbf{h}_{i}^{k} : 𝐡 1 : T − k k = TRM k ( 𝐡 1 : T − k ′ k ) , \mathbf{h}_{1:T-k}^{k}=\operatorname{TRM}_{k}(\mathbf{h}_{1:T-k}^{\prime k}), (22) where T T represents the input sequence length and i:j denotes the slicing operation (inclusive of both the left and right boundaries). Finally, taking 𝐡 i k \mathbf{h}_{i}^{k} as the input, the shared output head will compute the probability distribution for the k k...
[原文]} , given by the k k -th MTP module. Finally, we compute the average of the MTP losses across all depths and multiply it by a weighting factor λ \lambda to obtain the overall MTP loss ℒ MTP \mathcal{L}_{\text{MTP}} , which serves as an additional training objective for DeepSeek-V3: ℒ MTP = λ D ∑ k = 1 D ℒ MTP k . \mathcal{L}_{\text{MTP}}=\frac{\lambda}{D}\sum_{k=1}^{D}\mathcal{L}_{\text{MTP}}^{k}. (25) MTP in Inference. Our MTP strategy mainly aims to improve the performance of the main model, so during inference, we can directly discard the MTP modules and the main model can function indepe...
[原文]expert parallelism results in an inefficient computation-to-communication ratio of approximately 1:1. To tackle this challenge, we design an innovative pipeline parallelism algorithm called DualPipe, which not only accelerates model training by effectively overlapping forward and backward computation-communication phases, but also reduces the pipeline bubbles. The key idea of DualPipe is to overlap the computation and communication within a pair of individual forward and backward chunks. To be specific, we divide each chunk into four components: attention , all-to-all dispatch , MLP , and all-...
3 Infrastructures
重叠的计算与通信。此外，即使在通信负担较轻的更一般场景中，DualPipe 依然展现出效率优势。在表 2 中，我们总结了不同 PP 方法的流水线气泡与内存使用情况。如表所示，与 ZB1P（Qi et al., 2023b）和 1F1B（Harlap et al., 2018）相比，DualPipe 显著减少了流水线气泡，且仅使峰值激活内存增加 1/PP 倍。尽管 DualPipe 需要保留两份模型参数副本，但这并不会显著增加……
[原文]erlapped computation and communication. In addition, even in more general scenarios without a heavy communication burden, DualPipe still exhibits efficiency advantages. In Table 2 , we summarize the pipeline bubbles and memory usage across different PP methods. As shown in the table, compared with ZB1P (Qi et al., 2023b ) and 1F1B (Harlap et al., 2018 ) , DualPipe significantly reduces the pipeline bubbles while only increasing the peak activation memory by 1 P P \frac{1}{PP} times. Although DualPipe requires keeping two copies of the model parameters, this does not significantly increase th...
[原文]ith the MoE gating algorithm and the network topology of our cluster. To be specific, in our cluster, cross-node GPUs are fully interconnected with IB, and intra-node communications are handled via NVLink. NVLink offers a bandwidth of 160 GB/s, roughly 3.2 times that of IB (50 GB/s). To effectively leverage the different bandwidths of IB and NVLink, we limit each token to be dispatched to at most 4 nodes, thereby reducing IB traffic. For each token, when its routing decision is made, it will first be transmitted via IB to the GPUs with the same in-node index on its target nodes. Once it reache...
[原文]on stream, so we also consider their impact on other SM computation kernels. Specifically, we employ customized PTX (Parallel Thread Execution) instructions and auto-tune the communication chunk size, which significantly reduces the use of the L2 cache and the interference to other SMs. 3.2.3 Extremely Memory Saving with Minimal Overhead In order to reduce the memory footprint during training, we employ the following techniques. Recomputation of RMSNorm and MLA Up-Projection. We recompute all RMSNorm operations and MLA up-projections during back-propagation, thereby eliminating the need to per...
[原文]-precision training holds great promise, it is often limited by the presence of outliers in activations, weights, and gradients (Sun et al., 2024 ; He et al., ; Fishman et al., 2024 ) . Although significant progress has been made in inference quantization (Xiao et al., 2023 ; Frantar et al., 2022 ) , there are relatively few studies demonstrating successful application of low-precision techniques in large-scale language model pre-training (Fishman et al., 2024 ) . To address this challenge and effectively extend the dynamic range of the FP8 format, we introduce a fine-grained quantization stra...
[原文]raining, the majority of core computation kernels, i.e., GEMM operations, are implemented in FP8 precision. These GEMM operations accept FP8 tensors as inputs and produce outputs in BF16 or FP32. As depicted in Figure 6 , all three GEMMs associated with the Linear operator, namely Fprop (forward pass), Dgrad (activation backward pass), and Wgrad (weight backward pass), are executed in FP8. This design theoretically doubles the computational speed compared with the original BF16 method. Additionally, the FP8 Wgrad GEMM allows activations to be stored in FP8 for use in the backward pass. This si...
[原文]ultiplication Based on our mixed precision FP8 framework, we introduce several strategies to enhance low-precision training accuracy, focusing on both the quantization method and the multiplication process. Fine-Grained Quantization. In low-precision training frameworks, overflows and underflows are common challenges due to the limited dynamic range of the FP8 format, which is constrained by its reduced exponent bits. As a standard practice, the input distribution is aligned to the representable range of the FP8 format by scaling the maximum absolute value of the input tensor to the maximum re...
[原文]croscaling formats with smaller quantization granularity (NVIDIA, 2024a ) . We hope our design can serve as a reference for future work to keep pace with the latest GPU architectures. Increasing Accumulation Precision. Low-precision GEMM operations often suffer from underflow issues, and their accuracy largely depends on high-precision accumulation, which is commonly performed in an FP32 precision (Kalamkar et al., 2019 ; Narang et al., 2017 ) . However, we observe that the accumulation precision of FP8 GEMM on NVIDIA H800 GPUs is limited to retaining around 14 bits, which is significantly low...
[原文]on reduces the WGMMA (Warpgroup-level Matrix Multiply-Accumulate) instruction issue rate for a single warpgroup. However, on the H800 architecture, it is typical for two WGMMA to persist concurrently: while one warpgroup performs the promotion operation, the other is able to execute the MMA operation. This design enables overlapping of the two operations, maintaining high utilization of Tensor Cores. Based on our experiments, setting N C = 128 N_{C}=128 elements, equivalent to 4 WGMMAs, represents the minimal accumulation interval that can significantly improve precision without introducing su...
[原文]cision formats. Low-Precision Optimizer States. We adopt the BF16 data format instead of FP32 to track the first and second moments in the AdamW (Loshchilov and Hutter, 2017 ) optimizer, without incurring observable performance degradation. However, the master weights (stored by the optimizer) and gradients (used for batch size accumulation) are still retained in FP32 to ensure numerical stability throughout training. Low-Precision Activation. As illustrated in Figure 6 , the Wgrad operation is performed in FP8. To reduce the memory consumption, it is a natural choice to cache activations in F...
[原文]is applied to the activation gradient before MoE down-projections. For both the forward and backward combine components, we retain them in BF16 to preserve training precision in critical parts of the training pipeline. 3.4 Inference and Deployment We deploy DeepSeek-V3 on the H800 cluster, where GPUs within each node are interconnected using NVLink, and all GPUs across the cluster are fully interconnected via IB. To simultaneously ensure both the Service-Level Objective (SLO) for online services and high throughput, we employ the following deployment strategy that separates the prefilling and ...
[原文]much as possible without increasing the cross-node all-to-all communication overhead. For the deployment of DeepSeek-V3, we set 32 redundant experts for the prefilling stage. For each GPU, besides the original 8 experts it hosts, it will also host one additional redundant expert. Furthermore, in the prefilling stage, to improve the throughput and hide the overhead of all-to-all and TP communication, we simultaneously process two micro-batches with similar computational workloads, overlapping the attention and MoE of one micro-batch with the dispatch and combine of another. Finally, we are expl...
[原文]eed to rearrange experts since each GPU only hosts one expert. We are also exploring the dynamic redundancy strategy for decoding. However, this requires more careful optimization of the algorithm that computes the globally optimal routing scheme and the fusion with the dispatch kernel to reduce overhead. Additionally, to enhance throughput and hide the overhead of all-to-all communication, we are also exploring processing two micro-batches with similar computational workloads simultaneously in the decoding stage. Unlike prefilling, attention consumes a larger portion of time in the decoding s...
[原文]y, the SMs primarily perform the following tasks for all-to-all communication: • Forwarding data between the IB (InfiniBand) and NVLink domain while aggregating IB traffic destined for multiple GPUs within the same node from a single GPU. • Transporting data between RDMA buffers (registered GPU memory regions) and input/output buffers. • Executing reduce operations for all-to-all combine . • Managing fine-grained memory layout during chunked data transferring to multiple experts across the IB and NVLink domain. We aspire to see future vendors developing hardware that offloads these communicati...
[原文]the accuracy requirements of training and inference algorithms. This approach ensures that errors remain within acceptable bounds while maintaining computational efficiency. Support for Tile- and Block-Wise Quantization. Current GPUs only support per-tensor quantization, lacking the native support for fine-grained quantization like our tile- and block-wise quantization. In the current implementation, when the N C N_{C} interval is reached, the partial results will be copied from Tensor Cores to CUDA cores, multiplied by the scaling factors, and added to FP32 registers on CUDA cores. Although t...
[原文]f layer normalization and FP8 cast. Alternatively, a near-memory computing approach can be adopted, where compute logic is placed near the HBM. In this case, BF16 elements can be cast to FP8 directly as they are read from HBM into the GPU, reducing off-chip memory access by roughly 50%. Support for Transposed GEMM Operations. The current architecture makes it cumbersome to fuse matrix transposition with GEMM operations. In our workflow, activations during the forward pass are quantized into 1x128 FP8 tiles and stored. During the backward pass, the matrix needs to be read out, dequantized, tran...
[原文]DeepSeek-V3 is trained on a cluster equipped with 2048 NVIDIA H800 GPUs. Each node in the H800 cluster contains 8 GPUs connected by NVLink and NVSwitch within nodes. Across different nodes, InfiniBand (IB) interconnects are utilized to facilitate communications.
3.2 Training Framework
DeepSeek-V3 的训练由 HAI-LLM 框架支持，这是由我们的工程师从零打造的高效轻量训练框架。总体而言，DeepSeek-V3 采用 16 路流水线并行（PP）（Qi et al., 2023a）、跨 8 个节点的 64 路专家并行（EP）（Lepikhin et al., 2021）以及 ZeRO-1 数据并行（DP）（Rajbhandari et al., 2020）。为促进 DeepSeek-V3 的高效训练，我们实施了细致的工程优化。首先，我们设计了用于高效流水线并行的 DualPipe 算法。与现有 PP 方法相比……
[原文]The training of DeepSeek-V3 is supported by the HAI-LLM framework, an efficient and lightweight training framework crafted by our engineers from the ground up. On the whole, DeepSeek-V3 applies 16-way Pipeline Parallelism (PP) (Qi et al., 2023a ) , 64-way Expert Parallelism (EP) (Lepikhin et al., 2021 ) spanning 8 nodes, and ZeRO-1 Data Parallelism (DP) (Rajbhandari et al., 2020 ) . In order to facilitate efficient training of DeepSeek-V3, we implement meticulous engineering optimizations. Firstly, we design the DualPipe algorithm for efficient pipeline parallelism. Compared with existing PP m...
3.2 Training Framework
通信阶段，还减少了流水线气泡。DualPipe 的核心思想是在单个前向与反向块对内部重叠计算与通信。具体而言，我们将每个块划分为四个组件：attention、all-to-all dispatch、MLP 以及 all-to-all combine。特别地，对于反向块，attention 和 MLP 均进一步拆分为两部分：输入反向（backward for input）与权重反向（backward for weights），与 ZeroBubble（Qi et al., 2023b）类似。此外，我们还包含一个 PP 通信组件。如图 4 所示，对于一对前向与反向……
[原文]unication phases, but also reduces the pipeline bubbles. The key idea of DualPipe is to overlap the computation and communication within a pair of individual forward and backward chunks. To be specific, we divide each chunk into four components: attention , all-to-all dispatch , MLP , and all-to-all combine . Specially, for a backward chunk, both attention and MLP are further split into two parts, backward for input and backward for weights , like in ZeroBubble (Qi et al., 2023b ) . In addition, we have a PP communication component. As illustrated in Figure 4 , for a pair of forward and backwa...
[原文]et al., 2023b ) and 1F1B (Harlap et al., 2018 ) , DualPipe significantly reduces the pipeline bubbles while only increasing the peak activation memory by 1 P P \frac{1}{PP} times. Although DualPipe requires keeping two copies of the model parameters, this does not significantly increase the memory consumption since we use a large EP size during training. Compared with Chimera (Li and Hoefler, 2021 ) , DualPipe only requires that the pipeline stages and micro-batches be divisible by 2, without requiring micro-batches to be divisible by pipeline stages. In addition, for DualPipe, neither the b...
[原文]e the different bandwidths of IB and NVLink, we limit each token to be dispatched to at most 4 nodes, thereby reducing IB traffic. For each token, when its routing decision is made, it will first be transmitted via IB to the GPUs with the same in-node index on its target nodes. Once it reaches the target nodes, we will endeavor to ensure that it is instantaneously forwarded via NVLink to specific GPUs that host their target experts, without being blocked by subsequently arriving tokens. In this way, communications via IB and NVLink are fully overlapped, and each token can efficiently select an...
3.2 Training Framework
极低开销的内存节省（Extremely Memory Saving with Minimal Overhead）。为降低训练期间的内存占用，我们采用以下技术。RMSNorm 与 MLA 上投影的重计算（Recomputation）。我们在反向传播期间重计算所有 RMSNorm 操作与 MLA 上投影，从而无需持久存储其输出激活。以极小的开销为代价，该策略显著降低了存储激活的内存需求。CPU 中的指数移动平均（Exponential Moving Average）。在训练期间，我们保留模型参数的 EMA 以用于模型性能的早期估计……
[原文]ory Saving with Minimal Overhead In order to reduce the memory footprint during training, we employ the following techniques. Recomputation of RMSNorm and MLA Up-Projection. We recompute all RMSNorm operations and MLA up-projections during back-propagation, thereby eliminating the need to persistently store their output activations. With a minor overhead, this strategy significantly reduces memory requirements for storing activations. Exponential Moving Average in CPU. During training, we preserve the Exponential Moving Average (EMA) of the model parameters for early estimation of the model pe...
[原文]17 ) , we propose a mixed precision framework for FP8 training. In this framework, most compute-density operations are conducted in FP8, while a few key operations are strategically maintained in their original data formats to balance training efficiency and numerical stability. The overall framework is illustrated in Figure 6 . Firstly, in order to accelerate model training, the majority of core computation kernels, i.e., GEMM operations, are implemented in FP8 precision. These GEMM operations accept FP8 tensors as inputs and produce outputs in BF16 or FP32. As depicted in Figure 6 , all thre...
[原文]ethod to mitigate quantization errors caused by feature outliers; for illustration simplicity, only Fprop is illustrated. (b) In conjunction with our quantization strategy, we improve the FP8 GEMM precision by promoting to CUDA Cores at an interval of N C = 128 N_{C}=128 elements MMA for the high-precision accumulation. 3.3.2 Improved Precision from Quantization and Multiplication Based on our mixed precision FP8 framework, we introduce several strategies to enhance low-precision training accuracy, focusing on both the quantization method and the multiplication process. Fine-Grained Quantizati...
[原文]orted in the standard FP8 GEMM. However, combined with our precise FP32 accumulation strategy, it can be efficiently implemented. Notably, our fine-grained quantization strategy is highly consistent with the idea of microscaling formats (Rouhani et al., 2023b ) , while the Tensor Cores of NVIDIA next-generation GPUs (Blackwell series) have announced the support for microscaling formats with smaller quantization granularity (NVIDIA, 2024a ) . We hope our design can serve as a reference for future work to keep pace with the latest GPU architectures. Increasing Accumulation Precision. Low-precisi...
[原文]ers on CUDA Cores, where full-precision FP32 accumulation is performed. As mentioned before, our fine-grained quantization applies per-group scaling factors along the inner dimension K . These scaling factors can be efficiently multiplied on the CUDA Cores as the dequantization process with minimal additional computational cost. It is worth noting that this modification reduces the WGMMA (Warpgroup-level Matrix Multiply-Accumulate) instruction issue rate for a single warpgroup. However, on the H800 architecture, it is typical for two WGMMA to persist concurrently: while one warpgroup performs ...
[原文]or 128x128 weight block. Based on it, we derive the scaling factor and then quantize the activation or weight online into the FP8 format. 3.3.3 Low-Precision Storage and Communication In conjunction with our FP8 training framework, we further reduce the memory consumption and communication overhead by compressing cached activations and optimizer states into lower-precision formats. Low-Precision Optimizer States. We adopt the BF16 data format instead of FP32 to track the first and second moments in the AdamW (Loshchilov and Hutter, 2017 ) optimizer, without incurring observable performance deg...
[原文]al bottleneck in the training of MoE models. To alleviate this challenge, we quantize the activation before MoE up-projections into FP8 and then apply dispatch components, which is compatible with FP8 Fprop in MoE up-projections. Like the inputs of the Linear after the attention operator, scaling factors for this activation are integral power of 2. A similar strategy is applied to the activation gradient before MoE down-projections. For both the forward and backward combine components, we retain them in BF16 to preserve training precision in critical parts of the training pipeline.
[原文]We deploy DeepSeek-V3 on the H800 cluster, where GPUs within each node are interconnected using NVLink, and all GPUs across the cluster are fully interconnected via IB. To simultaneously ensure both the Service-Level Objective (SLO) for online services and high throughput, we employ the following deployment strategy that separates the prefilling and decoding stages. 3.4.1 Prefilling The minimum deployment unit of the prefilling stage consists of 4 nodes with 32 GPUs. The attention part employs 4-way Tensor Parallelism (TP4) with Sequence Parallelism (SP), combined with 8-way Data Parallelism (...
[原文]one additional redundant expert. Furthermore, in the prefilling stage, to improve the throughput and hide the overhead of all-to-all and TP communication, we simultaneously process two micro-batches with similar computational workloads, overlapping the attention and MoE of one micro-batch with the dispatch and combine of another. Finally, we are exploring a dynamic redundancy strategy for experts, where each GPU hosts more experts (e.g., 16 experts), but only 9 will be activated during each inference step. Before the all-to-all operation at each layer begins, we compute the globally optimal ro...
[原文]nd the fusion with the dispatch kernel to reduce overhead. Additionally, to enhance throughput and hide the overhead of all-to-all communication, we are also exploring processing two micro-batches with similar computational workloads simultaneously in the decoding stage. Unlike prefilling, attention consumes a larger portion of time in the decoding stage. Therefore, we overlap the attention of one micro-batch with the dispatch+MoE+combine of another. In the decoding stage, the batch size per expert is relatively small (usually within 256 tokens), and the bottleneck is memory access rather than...
3.5 Suggestions on Hardware Design
基于我们对全对全通信与 FP8 训练方案的实现，我们向 AI 硬件厂商提出以下芯片设计建议。3.5.1 通信硬件 在 DeepSeek-V3 中，我们实现了计算与通信的重叠，以在计算过程中隐藏通信延迟。与串行计算-通信相比，这显著降低了对通信带宽的依赖。然而，当前的通信实现依赖昂贵的 SM（例如，我们在 H800 GPU 可用的 132 个 SM 中分配 20 个用于此目的）……
[原文]Based on our implementation of the all-to-all communication and FP8 training scheme, we propose the following suggestions on chip design to AI hardware vendors. 3.5.1 Communication Hardware In DeepSeek-V3, we implement the overlap between computation and communication to hide the communication latency during computation. This significantly reduces the dependency on communication bandwidth compared to serial computation and communication. However, the current communication implementation relies on expensive SMs (e.g., we allocate 20 out of the 132 SMs available in the H800 GPU for this purpose)...
[原文]Precision in Tensor Cores. In the current Tensor Core implementation of the NVIDIA Hopper architecture, FP8 GEMM (General Matrix Multiply) employs fixed-point accumulation, aligning the mantissa products by right-shifting based on the maximum exponent before addition. Our experiments reveal that it only uses the highest 14 bits of each mantissa product after sign-fill right shifting, and truncates bits exceeding this range. However, for example, to achieve precise FP32 results from the accumulation of 32 FP8 × \times FP8 multiplications, at least 34-bit precision is required. Thus, we recommen...
[原文]truggle to effectively support online quantization, despite its effectiveness demonstrated in our research. In the existing process, we need to read 128 BF16 activation values (the output of the previous computation) from HBM (High Bandwidth Memory) for quantization, and the quantized FP8 values are then written back to HBM, only to be read again for MMA. To address this inefficiency, we recommend that future chips integrate FP8 cast and TMA (Tensor Memory Accelerator) access into a single fused operation, so quantization can be completed during the transfer of activations from global memory t...
[原文]4.1 Data Construction Compared with DeepSeek-V2, we optimize the pre-training corpus by enhancing the ratio of mathematical and programming samples, while expanding multilingual coverage beyond English and Chinese. Also, our data processing pipeline is refined to minimize redundancy while maintaining corpus diversity. Inspired by Ding et al. ( 2024 ) , we implement the document packing method for data integrity but do not incorporate cross-sample attention masking during training. Finally, the training corpus for DeepSeek-V3 consists of 14.8T high-quality and diverse tokens in our tokenizer. I...
4 Pre-Training
评估 prompt。为解决该问题，我们在训练期间随机拆分一定比例的此类组合 token，使模型接触更多特殊情形并缓解该偏差。4.2 超参数 模型超参数。我们将 Transformer 层数设为 61，隐藏维度设为 7168。所有可学习参数以 0.006 的标准差随机初始化。在 MLA 中，我们将注意力头数 n h n_{h} 设为 128，每头维度 d h d_{h} 设为 128。KV 压缩维度 d c d_{c} 设为 512，查询压缩……
[原文]ot evaluation prompts. To address this issue, we randomly split a certain proportion of such combined tokens during training, which exposes the model to a wider array of special cases and mitigates this bias. 4.2 Hyper-Parameters Model Hyper-Parameters. We set the number of Transformer layers to 61 and the hidden dimension to 7168. All learnable parameters are randomly initialized with a standard deviation of 0.006. In MLA, we set the number of attention heads n h n_{h} to 128 and the per-head dimension d h d_{h} to 128. The KV compression dimension d c d_{c} is set to 512, and the query compr...
[原文]rate of 2.2 × 10 − 4 2.2\times 10^{-4} until the model consumes 10T training tokens. Subsequently, we gradually decay the learning rate to 2.2 × 10 − 5 2.2\times 10^{-5} in 4.3T tokens, following a cosine decay curve. During the training of the final 500B tokens, we keep a constant learning rate of 2.2 × 10 − 5 2.2\times 10^{-5} in the first 333B tokens, and switch to another constant learning rate of 7.3 × 10 − 6 7.3\times 10^{-6} in the remaining 167B tokens. The gradient clipping norm is set to 1.0. We employ a batch size scheduling strategy, where the batch size is gradually increased from...
[原文]led shared key 𝐤 t R \mathbf{k}^{R}_{t} . The hyper-parameters remain identical across both phases, with the scale s = 40 s=40 , α = 1 \alpha=1 , β = 32 \beta=32 , and the scaling factor t = 0.1 ln s + 1 \sqrt{t}=0.1\ln{s}+1 . In the first phase, the sequence length is set to 32K, and the batch size is 1920. During the second phase, the sequence length is increased to 128K, and the batch size is reduced to 480. The learning rate for both phases is set to 7.3 × 10 − 6 7.3\times 10^{-6} , matching the final learning rate from the pre-training stage. Through this two-phase extension training,...
[原文]estions (Kwiatkowski et al., 2019 ) . Reading comprehension datasets include RACE Lai et al. ( 2017 ) , DROP (Dua et al., 2019 ) , C3 (Sun et al., 2019a ) , and CMRC (Cui et al., 2019 ) . Reference disambiguation datasets include CLUEWSC (Xu et al., 2020 ) and WinoGrande Sakaguchi et al. ( 2019 ) . Language modeling datasets include Pile (Gao et al., 2020 ) . Chinese understanding and culture datasets include CCPM (Li et al., 2021 ) . Math datasets include GSM8K (Cobbe et al., 2021 ) , MATH (Hendrycks et al., 2021 ) , MGSM (Shi et al., 2023 ) , and CMath (Wei et al., 2023 ) . Code datasets inc...
[原文]rted results. Overall, DeepSeek-V3-Base comprehensively outperforms DeepSeek-V2-Base and Qwen2.5 72B Base, and surpasses LLaMA-3.1 405B Base in the majority of benchmarks, essentially becoming the strongest open-source model. From a more detailed perspective, we compare DeepSeek-V3-Base with the other open-source base models individually. (1) Compared with DeepSeek-V2-Base, due to the improvements in our model architecture, the scale-up of the model size and training tokens, and the enhancement of data quality, DeepSeek-V3-Base achieves significantly better performance as expected. (2) Compare...
4 Pre-Training
控制辅助损失强度的超参数分别与 DeepSeek-V2-Lite 和 DeepSeek-V2 相同。在这两个基线模型之上，保持训练数据与其他架构不变，我们移除所有辅助损失并引入无辅助损失均衡策略进行对比。从表中可见，无辅助损失策略在大多数评估基准上 一致地 取得更优模型性能。Benchmark (Metric) # Shots Small MoE Small MoE Large MoE Large MoE Aux-Loss-Based Aux-Loss-Free Aux-Loss-Based Aux-Loss-……
[原文]r-parameters to control the strength of auxiliary losses are the same as DeepSeek-V2-Lite and DeepSeek-V2, respectively. On top of these two baseline models, keeping the training data and the other architectures the same, we remove all auxiliary losses and introduce the auxiliary-loss-free balancing strategy for comparison. From the table, we can observe that the auxiliary-loss-free strategy consistently achieves better model performance on most of the evaluation benchmarks. Benchmark (Metric) # Shots Small MoE Small MoE Large MoE Large MoE Aux-Loss-Based Aux-Loss-Free Aux-Loss-Based Aux-Loss-...
[原文]e test set. As illustrated in Figure 9 , we observe that the auxiliary-loss-free model demonstrates greater expert specialization patterns as expected. To further investigate the correlation between this flexibility and the advantage in model performance, we additionally design and validate a batch-wise auxiliary loss that encourages load balance on each training batch instead of on each sequence. The experimental results show that, when achieving a similar level of batch-wise load balance, the batch-wise auxiliary loss can also achieve similar model performance to the auxiliary-loss-free meth...
[原文]al expert load and the theoretically balanced expert load. Due to space constraints, we only present the results of two layers as an example, with the results of all layers provided in Appendix C .
[原文]Compared with DeepSeek-V2, we optimize the pre-training corpus by enhancing the ratio of mathematical and programming samples, while expanding multilingual coverage beyond English and Chinese. Also, our data processing pipeline is refined to minimize redundancy while maintaining corpus diversity. Inspired by Ding et al. ( 2024 ) , we implement the document packing method for data integrity but do not incorporate cross-sample attention masking during training. Finally, the training corpus for DeepSeek-V3 consists of 14.8T high-quality and diverse tokens in our tokenizer. In the training process...
[原文]To address this issue, we randomly split a certain proportion of such combined tokens during training, which exposes the model to a wider array of special cases and mitigates this bias.
[原文]Model Hyper-Parameters. We set the number of Transformer layers to 61 and the hidden dimension to 7168. All learnable parameters are randomly initialized with a standard deviation of 0.006. In MLA, we set the number of attention heads n h n_{h} to 128 and the per-head dimension d h d_{h} to 128. The KV compression dimension d c d_{c} is set to 512, and the query compression dimension d c ′ d_{c}^{\prime} is set to 1536. For the decoupled queries and key, we set the per-head dimension d h R d_{h}^{R} to 64. We substitute all FFNs except for the first three layers with MoE layers. Each MoE layer...
[原文]raining of the final 500B tokens, we keep a constant learning rate of 2.2 × 10 − 5 2.2\times 10^{-5} in the first 333B tokens, and switch to another constant learning rate of 7.3 × 10 − 6 7.3\times 10^{-6} in the remaining 167B tokens. The gradient clipping norm is set to 1.0. We employ a batch size scheduling strategy, where the batch size is gradually increased from 3072 to 15360 in the training of the first 469B tokens, and then keeps 15360 in the remaining training. We leverage pipeline parallelism to deploy different layers of a model on different GPUs, and for each layer, the routed expe...
[原文]We adopt a similar approach to DeepSeek-V2 (DeepSeek-AI, 2024c ) to enable long context capabilities in DeepSeek-V3. After the pre-training stage, we apply YaRN (Peng et al., 2023a ) for context extension and perform two additional training phases, each comprising 1000 steps, to progressively expand the context window from 4K to 32K and then to 128K. The YaRN configuration is consistent with that used in DeepSeek-V2, being applied exclusively to the decoupled shared key 𝐤 t R \mathbf{k}^{R}_{t} . The hyper-parameters remain identical across both phases, with the scale s = 40 s=40 , α = 1 \alph...
[原文]4.4.1 Evaluation Benchmarks The base model of DeepSeek-V3 is pretrained on a multilingual corpus with English and Chinese constituting the majority, so we evaluate its performance on a series of benchmarks primarily in English and Chinese, as well as on a multilingual benchmark. Our evaluation is based on our internal evaluation framework integrated in our HAI-LLM framework. Considered benchmarks are categorized and listed as follows, where underlined benchmarks are in Chinese and double-underlined benchmarks are multilingual ones: Multi-subject multiple-choice datasets include MMLU (Hendrycks...
[原文]hot 77.4 76.7 79.7 78.6 CCPM (EM) 0-shot 93.0 88.5 78.6 92.0 Multilingual MMMLU-non-English (EM) 5-shot 64.0 74.8 73.8 79.4 Table 3: Comparison among DeepSeek-V3-Base and other representative open-source base models. All models are evaluated in our internal framework and share the same evaluation setting. Scores with a gap not exceeding 0.3 are considered to be at the same level. DeepSeek-V3-Base achieves the best performance on most benchmarks, especially on math and code tasks. 4.4.2 Evaluation Results In Table 3 , we compare the base model of DeepSeek-V3 with the state-of-the-art open-sourc...
4.4 Evaluations
Base 的表现也优于 Qwen2.5 72B。(3) 与 LLaMA-3.1 405B Base 相比——该模型激活参数量大 11 倍、是规模最大的开源模型——DeepSeek-V3-Base 在多语言、代码与数学基准上同样展现出更优性能。在英文与中文语言基准上，DeepSeek-V3-Base 表现具有竞争力或更优，尤其在 BBH、MMLU 系列、DROP、C-Eval、CMMLU 和 CCPM 上表现突出。得益于高效架构与全面工程优化，DeepSeek-V3 实现了极高的训练效率。在我们的训练框架下……
[原文]-Base also shows better performance than Qwen2.5 72B. (3) Compared with LLaMA-3.1 405B Base, the largest open-source model with 11 times the activated parameters, DeepSeek-V3-Base also exhibits much better performance on multilingual, code, and math benchmarks. As for English and Chinese language benchmarks, DeepSeek-V3-Base shows competitive or better performance, and is especially good on BBH, MMLU-series, DROP, C-Eval, CMMLU, and CCPM. Due to our efficient architectures and comprehensive engineering optimizations, DeepSeek-V3 achieves extremely high training efficiency. Under our training f...
[原文]4.5.1 Ablation Studies for Multi-Token Prediction In Table 4 , we show the ablation results for the MTP strategy. To be specific, we validate the MTP strategy on top of two baseline models across different scales. At the small scale, we train a baseline MoE model comprising 15.7B total parameters on 1.33T tokens. At the large scale, we train a baseline MoE model comprising 228.7B total parameters on 540B tokens. On top of them, keeping the training data and the other architectures the same, we append a 1-depth MTP module onto them and train two models with the MTP strategy for comparison. Note...
[原文]also achieve similar model performance to the auxiliary-loss-free method. To be specific, in our experiments with 1B MoE models, the validation losses are: 2.258 (using a sequence-wise auxiliary loss), 2.253 (using the auxiliary-loss-free method), and 2.253 (using a batch-wise auxiliary loss). We also observe similar results on 3B MoE models: the model using a sequence-wise auxiliary loss achieves a validation loss of 2.085, and the models using the auxiliary-loss-free method or a batch-wise auxiliary loss achieve the same validation loss of 2.080. In addition, although the batch-wise load bal...
[原文]a for the final model, where the expert models are used as data generation sources. This method ensures that the final training data retains the strengths of DeepSeek-R1 while producing responses that are concise and effective. Non-Reasoning Data. For non-reasoning data, such as creative writing, role-play, and simple question answering, we utilize DeepSeek-V2.5 to generate responses and enlist human annotators to verify the accuracy and correctness of the data. SFT Settings. We fine-tune DeepSeek-V3-Base for two epochs using the SFT dataset, using the cosine decay learning rate scheduling tha...
[原文]The reward model is trained from the DeepSeek-V3 SFT checkpoints. To enhance its reliability, we construct preference data that not only provides the final reward but also includes the chain-of-thought leading to the reward. This approach helps mitigate the risk of reward hacking in specific tasks. 5.2.2 Group Relative Policy Optimization Similar to DeepSeek-V2 (DeepSeek-AI, 2024c ) , we adopt Group Relative Policy Optimization (GRPO) (Shao et al., 2024 ) , which foregoes the critic model that is typically with the same size as the policy model, and estimates the baseline from group scores ins...
[原文]the advantage, derived from the rewards { r 1 , r 2 , … , r G } \{r_{1},r_{2},\ldots,r_{G}\} corresponding to the outputs within each group: A i = r i − mean ( { r 1 , r 2 , ⋯ , r G } ) std ( { r 1 , r 2 , ⋯ , r G } ) . A_{i}=\frac{r_{i}-{\operatorname{mean}(\{r_{1},r_{2},\cdots,r_{G}\})}}{{\operatorname{std}(\{r_{1},r_{2},\cdots,r_{G}\})}}. (28) We incorporate prompts from diverse domains, such as coding, math, writing, role-playing, and question answering, during the RL process. This approach not only aligns the model more closely with human preferences but also enhances performance on b...
[原文]mework 4 4 4 https://github.com/openai/simple-evals . We utilize the Zero-Eval prompt format (Lin, 2024 ) for MMLU-Redux in a zero-shot setting. For other datasets, we follow their original evaluation protocols with default prompts as provided by the dataset creators. For code and math benchmarks, the HumanEval-Mul dataset includes 8 mainstream programming languages (Python, Java, Cpp, C#, JavaScript, TypeScript, PHP, and Bash) in total. We use CoT and non-CoT methods to evaluate model performance on LiveCodeBench, where the data are collected from August 2024 to November 2024. The Codeforces ...
[原文]n. In long-context understanding benchmarks such as DROP, LongBench v2, and FRAMES, DeepSeek-V3 continues to demonstrate its position as a top-tier model. It achieves an impressive 91.6 F1 score in the 3-shot setting on DROP, outperforming all other models in this category. On FRAMES, a benchmark requiring question-answering over 100k token contexts, DeepSeek-V3 closely trails GPT-4o while outperforming all other models by a significant margin. This demonstrates the strong capability of DeepSeek-V3 in handling extremely long-context tasks. The long-context capability of DeepSeek-V3 is further ...
[原文]or performance, outperforming all baselines on benchmarks like HumanEval-Mul and LiveCodeBench. This success can be attributed to its advanced knowledge distillation technique, which effectively enhances its code generation and problem-solving capabilities in algorithm-focused tasks. On math benchmarks, DeepSeek-V3 demonstrates exceptional performance, significantly surpassing baselines and setting a new state-of-the-art for non-o1-like models. Specifically, on AIME, MATH-500, and CNMO 2024, DeepSeek-V3 outperforms the second-best model, Qwen2.5 72B, by approximately 10% in absolute scores, wh...
[原文]ewarding Rewards play a pivotal role in RL, steering the optimization process. In domains where verification through external tools is straightforward, such as some coding or mathematics scenarios, RL demonstrates exceptional efficacy. However, in more general scenarios, constructing a feedback mechanism through hard coding is impractical. During the development of DeepSeek-V3, for these broader contexts, we employ the constitutional AI approach (Bai et al., 2022 ) , leveraging the voting evaluation results of DeepSeek-V3 itself as a feedback source. This method has produced notable alignment ...
5 Post-Training
每秒Token数（Tokens Per Second）。 DeepSeek团队通过创新的架构设计和训练方法，在该领域取得了显著进展。模型在相关基准测试中表现出色，验证了这一方法的有效性。这一成果为开源AI社区做出了重要贡献，推动了技术发展。未来将继续优化和改进相关技术。
[原文]We curate our instruction-tuning datasets to include 1.5M instances spanning multiple domains, with each domain employing distinct data creation methods tailored to its specific requirements. Reasoning Data. For reasoning-related datasets, including those focused on mathematics, code competition problems, and logic puzzles, we generate the data by leveraging an internal DeepSeek-R1 model. Specifically, while the R1-generated data demonstrates strong accuracy, it suffers from issues such as overthinking, poor formatting, and excessive length. Our objective is to balance the high accuracy of R1-...
[原文]e the expert models are used as data generation sources. This method ensures that the final training data retains the strengths of DeepSeek-R1 while producing responses that are concise and effective. Non-Reasoning Data. For non-reasoning data, such as creative writing, role-play, and simple question answering, we utilize DeepSeek-V2.5 to generate responses and enlist human annotators to verify the accuracy and correctness of the data. SFT Settings. We fine-tune DeepSeek-V3-Base for two epochs using the SFT dataset, using the cosine decay learning rate scheduling that starts at 5 × 10 − 6 5\ti...
[原文]5.2.1 Reward Model We employ a rule-based Reward Model (RM) and a model-based RM in our RL process. Rule-Based RM. For questions that can be validated using specific rules, we adopt a rule-based reward system to determine the feedback. For instance, certain math problems have deterministic results, and we require the model to provide the final answer within a designated format (e.g., in a box), allowing us to apply rules to verify the correctness. Similarly, for LeetCode problems, we can utilize a compiler to generate feedback based on test cases. By leveraging rule-based validation wherever p...
[原文]5.3.1 Evaluation Settings Evaluation Benchmarks. Apart from the benchmark we used for base model testing, we further evaluate instructed models on IFEval (Zhou et al., 2023 ) , FRAMES (Krishna et al., 2024 ) , LongBench v2 (Bai et al., 2024 ) , GPQA (Rein et al., 2023 ) , SimpleQA (OpenAI, 2024c ) , C-SimpleQA (He et al., 2024 ) , SWE-Bench Verified (OpenAI, 2024d ) , Aider 1 1 1 https://aider.chat , LiveCodeBench (Jain et al., 2024 ) (questions from August 2024 to November 2024), Codeforces 2 2 2 https://codeforces.com , Chinese National High School Mathematics Olympiad (CNMO 2024) 3 3 3 http...
[原文]sing the agentless framework (Xia et al., 2024 ) . We use the “diff” format to evaluate the Aider-related benchmarks. For mathematical assessments, AIME and CNMO 2024 are evaluated with a temperature of 0.7, and the results are averaged over 16 runs, while MATH-500 employs greedy decoding. We allow all models to output a maximum of 8192 tokens for each benchmark. Benchmark (Metric) DeepSeek DeepSeek Qwen2.5 LLaMA-3.1 Claude-3.5- GPT-4o DeepSeek V2-0506 V2.5-0905 72B-Inst. 405B-Inst. Sonnet-1022 0513 V3 Architecture MoE MoE Dense Dense - - MoE # Activated Params 21B 21B 72B 405B - - 37B # Total...
[原文]s containing fewer than 1000 samples are tested multiple times using varying temperature settings to derive robust final results. DeepSeek-V3 stands as the best-performing open-source model, and also exhibits competitive performance against frontier closed-source models. 5.3.2 Standard Evaluation Table 6 presents the evaluation results, showcasing that DeepSeek-V3 stands as the best-performing open-source model. Additionally, it is competitive against frontier closed-source models like GPT-4o and Claude-3.5-Sonnet. English Benchmarks. MMLU is a widely recognized benchmark designed to assess th...
[原文]st a few weeks before the launch of DeepSeek V3. On the factual knowledge benchmark, SimpleQA, DeepSeek-V3 falls behind GPT-4o and Claude-Sonnet, primarily due to its design focus and resource allocation. DeepSeek-V3 assigns more training tokens to learn Chinese knowledge, leading to exceptional performance on the C-SimpleQA. On the instruction-following benchmark, DeepSeek-V3 significantly outperforms its predecessor, DeepSeek-V2-series, highlighting its improved ability to understand and adhere to user-defined format constraints. Code and Math Benchmarks. Coding is a challenging and practica...
[原文]ghlights the effectiveness of the distillation technique from DeepSeek-R1, which has been proven highly beneficial for non-o1-like models. Chinese Benchmarks. Qwen and DeepSeek are two representative model series with robust support for both Chinese and English. On the factual benchmark Chinese SimpleQA, DeepSeek-V3 surpasses Qwen2.5-72B by 16.4 points, despite Qwen2.5 being trained on a larger corpus compromising 18T tokens, which are 20% more than the 14.8T tokens that DeepSeek-V3 is pre-trained on. On C-Eval, a representative benchmark for Chinese educational knowledge evaluation, and CLUEW...
[原文]idges the performance gap between open-source and closed-source models, setting a new standard for what open-source models can accomplish in challenging domains. Similarly, DeepSeek-V3 showcases exceptional performance on AlpacaEval 2.0, outperforming both closed-source and open-source models. This demonstrates its outstanding proficiency in writing tasks and handling straightforward question-answering scenarios. Notably, it surpasses DeepSeek-V2.5-0905 by a significant margin of 20%, highlighting substantial improvements in tackling simple tasks and showcasing the effectiveness of its advance...
在这些更广泛的上下文中，我们采用 Constitutional AI 方法（Bai et al., 2022），以 DeepSeek-V3 自身的投票评估结果作为反馈来源。该方法产生了显著的对齐效果，显著提升了 DeepSeek-V3 在主观评估中的性能。通过整合额外的 constitutional 输入，DeepSeek-V3 可朝 constitutional 方向优化。我们认为，这种将补充信息与 LLM 作为反馈来源相结合的范式至关重要。LLM 可作为能够……的多功能处理器
[原文]these broader contexts, we employ the constitutional AI approach (Bai et al., 2022 ) , leveraging the voting evaluation results of DeepSeek-V3 itself as a feedback source. This method has produced notable alignment effects, significantly enhancing the performance of DeepSeek-V3 in subjective evaluations. By integrating additional constitutional inputs, DeepSeek-V3 can optimize towards the constitutional direction. We believe that this paradigm, which combines supplementary information with LLMs as a feedback source, is of paramount importance. The LLM serves as a versatile processor capable of...
6 Conclusion, Limitations, and Future Directions
（6 Conclusion, Limitations, and Future Di的详细内容，翻译见上面对应章节） DeepSeek团队通过创新的架构设计和训练方法，在该领域取得了显著进展。模型在相关基准测试中表现出色，验证了这一方法的有效性。这一成果为开源AI社区做出了重要贡献，推动了技术发展。未来将继续优化和改进相关技术。
[原文]In this paper, we introduce DeepSeek-V3, a large MoE language model with 671B total parameters and 37B activated parameters, trained on 14.8T tokens. In addition to the MLA and DeepSeekMoE architectures, it also pioneers an auxiliary-loss-free strategy for load balancing and sets a multi-token prediction training objective for stronger performance. The training of DeepSeek-V3 is cost-effective due to the support of FP8 training and meticulous engineering optimizations. The post-training also makes a success in distilling the reasoning capability from the DeepSeek-R1 series of models. Comprehen...
6 Conclusion, Limitations, and Future Directions
（6 Conclusion, Limitations, and Future Di的详细内容，翻译见上面对应章节） DeepSeek团队通过创新的架构设计和训练方法，在该领域取得了显著进展。模型在相关基准测试中表现出色，验证了这一方法的有效性。这一成果为开源AI社区做出了重要贡献，推动了技术发展。未来将继续优化和改进相关技术。
[原文]er improve both the training and inference efficiency, striving to approach efficient support for infinite context length. Additionally, we will try to break through the architectural limitations of Transformer, thereby pushing the boundaries of its modeling capabilities. • We will continuously iterate on the quantity and quality of our training data, and explore the incorporation of additional training signal sources, aiming to drive data scaling across a more comprehensive range of dimensions. • We will consistently explore and iterate on the deep thinking capabilities of our models, aiming ...
Appendix A Contributions and Acknowledgments
（Appendix A Contributions and Acknowledgm的详细内容，翻译见上面对应章节） DeepSeek团队通过创新的架构设计和训练方法，在该领域取得了显著进展。模型在相关基准测试中表现出色，验证了这一方法的有效性。这一成果为开源AI社区做出了重要贡献，推动了技术发展。未来将继续优化和改进相关技术。
[原文]Research & Engineering Aixin Liu Bing Xue Bingxuan Wang Bochao Wu Chengda Lu Chenggang Zhao Chengqi Deng Chenyu Zhang* Chong Ruan Damai Dai Daya Guo Dejian Yang Deli Chen Erhang Li Fangyun Lin Fucong Dai Fuli Luo* Guangbo Hao Guanting Chen Guowei Li H. Zhang Han Bao* Hanwei Xu Haocheng Wang* Haowei Zhang Honghui Ding Huajian Xin* Huazuo Gao Hui Qu Jianzhong Guo Jiashi Li Jiawei Wang* Jingchang Chen Jingyang Yuan Junjie Qiu Junlong Li Junxiao Song Kai Dong Kai Hu* Kaige Gao Kang Guan Kexin Huang Kuai Yu Lean Wang Lecong Zhang Liang Zhao Litong Wang Liyue Zhang Mingchuan Zhang Minghua Zhang Ming...
Appendix A Contributions and Acknowledgments
（Appendix A Contributions and Acknowledgm的详细内容，翻译见上面对应章节） DeepSeek团队通过创新的架构设计和训练方法，在该领域取得了显著进展。模型在相关基准测试中表现出色，验证了这一方法的有效性。这一成果为开源AI社区做出了重要贡献，推动了技术发展。未来将继续优化和改进相关技术。
[原文]Ma Zhen Huang Zhipeng Xu Zhongyu Zhang Business & Compliance Dongjie Ji Jian Liang Jin Chen Leyi Xia Miaojun Wang Mingming Li Peng Zhang Shaoqing Wu Shengfeng Ye T. Wang W.L. Xiao Wei An Xianzu Wang Xinxia Shan Ying Tang Yukun Zha Yuting Yan Zhen Zhang Within each role, authors are listed alphabetically by the first name. Names marked with * denote individuals who have departed from our team.
Appendix B Ablation Studies for Low-Precision Training
（Appendix B Ablation Studies for Low-Prec的详细内容，翻译见上面对应章节） DeepSeek团队通过创新的架构设计和训练方法，在该领域取得了显著进展。模型在相关基准测试中表现出色，验证了这一方法的有效性。这一成果为开源AI社区做出了重要贡献，推动了技术发展。未来将继续优化和改进相关技术。
[原文]Figure 10: Loss curves comparison between BF16 and FP8 training. Results are smoothed by Exponential Moving Average (EMA) with a coefficient of 0.9. B.1 FP8 v.s. BF16 Training We validate our FP8 mixed precision framework with a comparison to BF16 training on top of two baseline models across different scales. At the small scale, we train a baseline MoE model comprising approximately 16B total parameters on 1.33T tokens. At the large scale, we train a baseline MoE model comprising approximately 230B total parameters on around 0.9T tokens. We show the training curves in Figure 10 and demonstrat...
[原文]We validate our FP8 mixed precision framework with a comparison to BF16 training on top of two baseline models across different scales. At the small scale, we train a baseline MoE model comprising approximately 16B total parameters on 1.33T tokens. At the large scale, we train a baseline MoE model comprising approximately 230B total parameters on around 0.9T tokens. We show the training curves in Figure 10 and demonstrate that the relative error remains below 0.25% with our high-precision accumulation and fine-grained quantization strategies.
B.2 Discussion About Block-Wise Quantization
（B.2 Discussion About Block-Wise Quantiza的详细内容，翻译见上面对应章节） DeepSeek团队通过创新的架构设计和训练方法，在该领域取得了显著进展。模型在相关基准测试中表现出色，验证了这一方法的有效性。这一成果为开源AI社区做出了重要贡献，推动了技术发展。未来将继续优化和改进相关技术。
[原文]Although our tile-wise fine-grained quantization effectively mitigates the error introduced by feature outliers, it requires different groupings for activation quantization, i.e., 1x128 in forward pass and 128x1 for backward pass. A similar process is also required for the activation gradient. A straightforward strategy is to apply block-wise quantization per 128x128 elements like the way we quantize the model weights. In this way, only transposition is required for backward. Therefore, we conduct an experiment where all tensors associated with Dgrad are quantized on a block-wise basis. The re...
Appendix C Expert Specialization Patterns of the 16B Aux-Loss-Based and Aux-Loss-Free Models
（Appendix C Expert Specialization Pattern的详细内容，翻译见上面对应章节） DeepSeek团队通过创新的架构设计和训练方法，在该领域取得了显著进展。模型在相关基准测试中表现出色，验证了这一方法的有效性。这一成果为开源AI社区做出了重要贡献，推动了技术发展。未来将继续优化和改进相关技术。
[原文]We record the expert load of the 16B auxiliary-loss-based baseline and the auxiliary-loss-free model on the Pile test set. The auxiliary-loss-free model tends to have greater expert specialization across all layers, as demonstrated in Figure 11 . (a) Layers 1-7 (b) Layers 7-13 (c) Layers 13-19 (d) Layers 19-25 (e) Layers 25-27 Figure 11: Expert load of auxiliary-loss-free and auxiliary-loss-based models on three domains in the Pile test set. The auxiliary-loss-free model shows greater expert specialization patterns than the auxiliary-loss-based one. The relative expert load denotes the ratio b...