arXiv subjects

Donglei Wu

Publications and source records attributed to Donglei Wu.

3 recordsLinked to original sources

arXiv2026

High-accuracy Low-Bit KV-Cache Quantization via Local Distribution Restoration

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.

cs.LG↗

arXiv2026

TTKV: Temporal-Tiered KV Cache for Long-Context LLM Inference

Key-value (KV) caching is critical for efficient inference in large language models (LLMs), yet its memory footprint scales linearly with context length, resulting in a severe scalability bottleneck. Existing approaches largely treat KV states as equally important across time, implicitly assuming uniform precision and accessibility. However, this assumption contrasts with human memory systems, where memories vary in clarity, recall frequency, and relevance with temporal proximity.Motivated by this insight, we propose TTKV, a KV cache management framework that maps the human memory system onto the KV cache. TTKV partitions the KV cache into temporal tiers with heterogeneous capacity and precision. The design addresses three aspects: (1) Tier Layout, decoupling fast and slow memory using HBM and DRAM; (2) Tier Content, assigning more recent KV states to faster, higher-precision tiers based on temporal proximity; and (3) Tier Interaction, employing block-wise streaming attention to overlap communication and computation when accessing slow tiers. Experiments show that TTKV reduces cross-tier traffic by 5.94x on 128K-context tasks, achieving up to 76% latency reduction and 2x throughput improvement over strong baselines.

cs.CL↗

arXiv2025

HybridEP: Scaling Expert Parallelism to Cross-Datacenter Scenario via Hybrid Expert/Data Transmission

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.

cs.DC↗