本周六下午 14:00-16:00，我们将在学堂 112【线下】给大家带来 Nathan Chen 和张宇的报告。报告和预训练与模型架构相关。
摘要
Residual connections with PreNorm are standard in modern LLMs, yet they accumulate all layer outputs with fixed unit weights. This uniform aggregation causes uncontrolled hidden-state growth with depth, progressively diluting each layer’s contribution. We propose Attention Residuals (AttnRes), which replaces this fixed accumulation with softmax attention over preceding layer outputs, allowing each layer to selectively aggregate earlier representations with learned, input-dependent weights. To address the memory and communication overhead of attending over all preceding layer outputs for large-scale model training, we introduce Block AttnRes, which partitions layers into blocks and attends over block-level representations, reducing the memory footprint while preserving most of the gains of full AttnRes. Combined with cache-based pipeline communication and a two-phase computation strategy, Block AttnRes becomes a practical drop-in replacement for standard residual connections with minimal overhead.
Scaling law experiments confirm that the improvement is consistent across model sizes, and ablations validate the benefit of content-dependent depth-wise selection. We further integrate AttnRes into the Kimi Linear architecture (48B total / 3B activated parameters) and pre-train on 1.4T tokens, where AttnRes mitigates PreNorm dilution, yielding more uniform output magnitudes and gradient distribution across depth, and improves downstream performance across all evaluated tasks.
讲者
Nathan Chen and Yu Zhang are researchers in the Scaling Team at Moonshot AI (Kimi), and two of the main authors of Attention Residuals. Nathan’s research interests include model architecture, efficient attention, and continual learning. Yu’s research focuses on efficient text generation models and hardware-efficient methods for sequence modeling, and he is the primary maintainer of Flash Linear Attention (FLA).
欢迎全体同学参加~
Make IIIS Great Again!
