Just now, DeepSeek V4 updates with DSpark, improving inference speed by 80%
DeepSeek has updated its DeepSeek V4 model with the DSpark speculative decoding framework, achieving a significant 60-85% speedup in generation for Flash models and 57-78% for Pro models while maintaining the same overall throughput.
This engineering-focused update, rather than a core architectural change, introduces DSpark to address latency and throughput bottlenecks in high-concurrency production environments. DSpark combines high-throughput parallel generation with adaptive load-aware verification. Its key innovations include a semi-autoregressive generation architecture to model dependencies within token blocks and a hardware-aware confidence-scheduled verification system. This system uses a confidence head to predict token acceptance probabilities, allowing it to dynamically optimize verification length per request and allocate compute only to tokens with the highest expected payoff.
The asynchronous scheduler is designed for real-world deployment, ensuring zero-overhead scheduling and continuous CUDA graph replay while preserving the target model's output distribution. In tests across mathematical reasoning, code generation, and daily dialogue, DSpark outperformed state-of-the-art models like Eagle3 and DFlash, increasing average acceptance length by 26.7%-30.9% and 16.3%-18.4% respectively on Qwen3 target models.
DeepSeek also open-sourced DeepSpec, a full-stack codebase for training and evaluating speculative decoding draft models, providing a standardized toolkit that includes data preparation tools, model implementations, training code, and evaluation scripts.
marsbitHace 6 hora(s)
