Deep Learning

By introducing DeepSeek Sparse Attention (DSA), a scalable reinforcement learning framework, and a large-scale agentic task synthesis pipeline, DeepSeek-V3.2 (DeepSeek-AI, 2025) achieves reasoning capabilities and agent performance comparable to GPT-5. Fig. 1. Benchmark of DeepSeek-V3.2 and its counterparts. (Image source: DeepSeek-AI, 2025) DeepSeek Sparse Attention Fig. 2. Attention architecture of DeepSeek-V3.2, where DSA is instantiated under MLA. (Image source: DeepSeek-AI, 2025) ...

From the evolution of the GPT series, researchers have gradually realized that as long as model parameters, training data, and compute resources are continuously scaled up, the performance of large models improves along a stable and predictable path. This predictability is characterized by Scaling Laws, which provide the theoretical foundation and practical confidence for high-cost pre-training. As model scale, alignment techniques, and inference-time compute co-evolve, the boundaries of AI capabilities are being systematically pushed. Scaling laws are not only the foundation for building next-generation models but also a key methodology for continuously improving model capabilities under compute constraints. ...

As the parameters of Large Language Models (LLMs) continue to grow, deploying and serving these models presents significant challenges. vLLM is an open-source library designed for fast, convenient, and cost-effective LLM inference and online serving. Its core lies in the PagedAttention algorithm, which efficiently manages the KV Cache in the attention mechanism. Evaluation Metrics To evaluate the performance of LLM inference and serving engines, we primarily focus on the following metrics: ...

DeepSeek AI successively released DeepSeek-V2 (DeepSeek-AI, 2024) and DeepSeek-V3 (DeepSeek-AI, 2024), two powerful Mixture-of-Experts (MoE) language models that significantly optimize training costs and inference efficiency while maintaining state-of-the-art performance. DeepSeek-V2 has a total of 236B parameters, activating 21B per token, while DeepSeek-V3 further expands to 671B total parameters, activating 37B per token. Both support a 128K context length. The core innovations of these two models lie in the adoption of Multi-head Latent Attention (MLA) and the DeepSeekMoE architecture (Dai et al., 2024). MLA drastically reduces GPU memory usage during inference by compressing the Key-Value (KV) cache into low-dimensional latent vectors, improving efficiency. DeepSeekMoE achieves stronger expert specialization capabilities and more economical training costs through fine-grained expert segmentation and shared expert isolation. Building upon V2, DeepSeek-V3 further introduces an Auxiliary-Loss-Free Load Balancing strategy (Wang et al., 2024) and the Multi-Token Prediction (MTP) (Gloeckle et al., 2024) training objective, further enhancing model performance and training efficiency. ...

Background Recently, the number of parameters in large models has been continuously increasing, from the initial billions to today’s hundreds of billions or even trillions. While large models have brought unprecedented application effects, they have also triggered a series of severe challenges in computing resources, memory management, and training stability. Therefore, this blog summarizes some commonly used distributed parallel training and memory management techniques, hoping to help everyone better train and optimize large models. ...