Yanqi Pan

Gradwell Dzikanyanga, Yanqi Pan, Weihao Yang et al.

Long-context large language model inference relies on the KV cache to avoid redundant attention computation, but incurs high memory and bandwidth overheads. Low-bit KV-cache quantization reduces this cost, yet it severely degrade quality; particularly, one-bit quantization reduces accuracy from 84.2% to 47.8% on Llama-3.1-8B under RULER. Rather than common beliefs that absolute error of logits, we find that the root cause is structured local misranking, where the distribution of logits in top-K region is drifted. We thereby propose local distribution restoration, a new technique that detects steps with high local distribution risk from quantized-logit features and restores only the selected top-K candidate distribution before token selection. We implement DGAP to achieve local distribution restoration, with efficient risk detcetors and correctors. Expeirments show that on Llama-3.1-8B, DGAP recovers K1V1 RULER accuracy from 47.8% to 83.2% and reduces distribution drift from 0.38 to 0.14; across Llama, Mistral, and Qwen models, it preserves the persistent low-bit KV-cache footprint with modest decode overhead.

Yanqi Pan, Qinghao Huang, Weihao Yang

We proudly introduce Memory-as-Asset, a new memory paradigm towards human-centric artificial general intelligence (AGI). In this paper, we formally emphasize that human-centric, personal memory management is a prerequisite for complementing the collective knowledge of existing large language models (LLMs) and extending their knowledge boundaries through self-evolution. We introduce three key features that shape the Memory-as-Asset era: (1) Memory in Hand, which emphasizes human-centric ownership to maximize benefits to humans; (2) Memory Group, which provides collaborative knowledge formation to avoid memory islands, and (3) Collective Memory Evolution, which enables continuous knowledge growth to extend the boundary of knowledge towards AGI. We finally give a potential three-layer memory infrastructure to facilitate the Memory-as-Asset paradigm, with fast personal memory storage, an intelligent evolution layer, and a decentralized memory exchange network. Together, these components outline a foundational architecture in which personal memories become persistent digital assets that can be accumulated, shared, and evolved over time. We believe this paradigm provides a promising path toward scalable, human-centric AGI systems that continuously grow through the collective experiences of individuals and intelligent agents.

Weihao Yang, Hao Huang, Donglei Wu et al.

Mixture-of-Experts (MoE) has become a popular architecture for scaling large models. However, the rapidly growing scale outpaces model training on a single DC, driving a shift toward a more flexible, cross-DC training paradigm. Under this, Expert Parallelism (EP) of MoE faces significant scalability issues due to the limited cross-DC bandwidth. Specifically, existing EP optimizations attempt to overlap data communication and computation, which has little benefit in low-bandwidth scenarios due to a much longer data communication time. Therefore, the trends of cross-DC EP scaling is fast becoming a critical roadblock to the continued growth of MoE models. To address this, we propose HybridEP, a modeling-guided framework to optimize EP under constrained bandwidth. Our key idea is to dynamically transform the spatial placement of experts to reduce data communication traffic and frequency, thereby minimizing EP's communication overheads. However, it is non-trivial to find the optimal solution because it complicates the original communication pattern by mixing data and expert communication. We therefore build a stream-based model to determine the optimal transmission ratio. Guided by this, we incorporate two techniques: (1) domain-based partition to construct the mapping between hybrid patterns and specific communication topology at GPU level, and (2) parameter-efficient migration to further refine this topology by reducing expert transmission overhead and enlarging the domain size. Combining all these designs, HybridEP can be considered as a more general EP with better scalability. Experimental results show that HybridEP outperforms existing state-of-the-art MoE training systems by up to 5.6x under constrained bandwidth. We further compare HybridEP and EP on large-scale simulations. HybridEP achieves up to 1.45x speedup with 1k DCs under different bandwidths.