[Paper Review] DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model

DeepSeek-AI, Liu, Aixin+155 more|arXiv (Cornell University)|May 7, 2024

Expert finding and Q&A systems97 citations

TL;DR

DeepSeek-V2 is a 236B parameter open-source MoE language model with 21B activated per token, 128K context, and novel MLA and DeepSeekMoE architectures that enable economical training and efficient inference, achieving top-tier open-source performance.

ABSTRACT

We present DeepSeek-V2, a strong Mixture-of-Experts (MoE) language model characterized by economical training and efficient inference. It comprises 236B total parameters, of which 21B are activated for each token, and supports a context length of 128K tokens. DeepSeek-V2 adopts innovative architectures including Multi-head Latent Attention (MLA) and DeepSeekMoE. MLA guarantees efficient inference through significantly compressing the Key-Value (KV) cache into a latent vector, while DeepSeekMoE enables training strong models at an economical cost through sparse computation. Compared with DeepSeek 67B, DeepSeek-V2 achieves significantly stronger performance, and meanwhile saves 42.5% of training costs, reduces the KV cache by 93.3%, and boosts the maximum generation throughput to 5.76 times. We pretrain DeepSeek-V2 on a high-quality and multi-source corpus consisting of 8.1T tokens, and further perform Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) to fully unlock its potential. Evaluation results show that, even with only 21B activated parameters, DeepSeek-V2 and its chat versions still achieve top-tier performance among open-source models.

Motivation & Objective

• Address the resource and efficiency challenges of large language models through economical training and fast inference.
• Develop architectures to reduce the KV cache and enable scalable MoE training.
• Achieve strong performance on English and Chinese benchmarks while reducing training costs and improving inference throughput.

Proposed method

• Introduce Multi-head Latent Attention (MLA) with low-rank key-value joint compression to reduce KV cache during inference.
• Adopt DeepSeekMoE for FFNs to enable training strong models at economical costs via sparse routing and fine-grained experts.
• Use decoupled Rotary Position Embedding to maintain RoPE compatibility with MLA.
• Implement device-limited routing, auxiliary load-balancing losses, and a token-dropping strategy to control communication and compute in MoE.
• Pretrain on an 8.1T multi-source corpus, followed by supervised fine-tuning (SFT) and reinforcement learning (RL) with Group Relative Policy Optimization (GRPO) to align the model.
• Extend context length to 128K using YaRN for long-context extension.

Experimental results

Research questions

• RQ1 How does MLA compare to standard MHA, GQA, and MQA in terms of performance and KV cache efficiency?
• RQ2 Can DeepSeekMoE enable strong model performance at lower training costs compared to dense equivalents or other MoE architectures?
• RQ3 How does DeepSeek-V2 perform on English and Chinese benchmarks relative to open-source baselines with similar activated parameter counts?
• RQ4 What is the impact of SFT and RL alignment on DeepSeek-V2 Chat performance across English and Chinese tasks?

Key findings

• DeepSeek-V2 achieves top-tier performance among open-source models with only 21B activated parameters.
• Compared with DeepSeek 67B, it saves 42.5% of training costs, reduces the KV cache by 93.3%, and increases maximum generation throughput by 5.76x.
• The model has 236B total parameters with 21B activated per token and supports a 128K context length.
• DeepSeek-V2 Chat (RL) achieves strong scores on AlpacaEval 2.0 (38.9 length-controlled win rate), MT-Bench (8.97), and AlignBench (7.91).
• In Chinese benchmarks, DeepSeek-V2 Chat (RL) outperforms open-source models and many closed-source models on AlignBench.
• DeepSeek-V2-Lite (15.7B total, 2.4B activated) is released for the community.