Tag
研究人员提出 ORL（On-Policy Residual Learning）与自研推理引擎 dlmrt，使扩散语言模型在单张 RTX 5090 上达到 2,788 tokens/s，8×B200 上达 7,708 tokens/s，端到端吞吐提升最高 6.9 倍且质量不降，模型、训练代码与推理引擎全部开源。
This paper proposes L0-MoE, a lightweight Mixture-of-Experts approach using L0-regularization to accelerate dense Large Language Models with up to 2.5x speedup while maintaining competitive performance.
Google Research introduces a new architecture using frozen Multi-Token Prediction to accelerate Gemini Nano models on Pixel devices, significantly improving speed and energy efficiency for on-device AI features.
JetFlow is a speculative decoding framework that breaks the scaling ceiling by combining one-forward drafting efficiency with branch-wise causal conditioning, achieving up to 9.64x speedup on math benchmarks and outperforming prior methods on dense and MoE Qwen3 models.