arXiv subjects

Zhaogeng Li

Publications and source records attributed to Zhaogeng Li.

3 recordsLinked to original sources

arXiv2026

Adapting Tree-Structured Speculative Decoding to DeepSeek-V4 for Efficient Inference

Repeated execution of the target model during autoregressive decoding is a major source of LLM inference latency. Unlike linear speculation, which follows a single candidate chain, tree-structured speculation retains multiple branches from shared prefixes; under the same budget, this broader coverage can improve acceptance and efficiency. Adapting it to DeepSeek-V4 is nontrivial: its CSA/HCA online compressed attention concentrates the difficulty on the target-verify side, where branches diverging from a shared prefix compress into different states, breaking cross-branch state consistency. We integrate tree-structured speculative decoding into the DeepSeek-V4-Flash pipeline via branch-aware causal verification, temporary state isolation, and accepted-path state refresh, keeping verification and compressed-state updates consistent across branches. Across budgets D=5 to D=8, batch sizes 1 to 64, and three datasets (GSM8K, MBPP, ShareGPT), tree speculation achieves a higher accepted length than the matched linear configurations in all settings (e.g., at D=8 about 2.83--3.41 versus 2.39--2.84) and improves throughput in nearly all configurations---marginal only at the smallest budget---by up to about 18.5%. More importantly, the gains follow stable, transferable regularities: the relative gain grows with the budget and is most pronounced for less predictable workloads at small-to-medium batch sizes, while beyond a certain budget throughput plateaus and decouples from the still-rising accepted length. These results show that retaining multiple candidate paths under the same budget can effectively improve DeepSeek-V4 decoding efficiency, and offer experience for adapting speculative decoding to future models with compressed, sparse, or structured context representations.

cs.CL↗

arXiv2026

Revealing the Challenges of Attention-FFN Disaggregation for Modern MoE Models and Hardware Systems

Deploying large-scale MoE models presents challenges in memory capacity and bandwidth for expert activation. While Attention-FFN Disaggregation (AFD) has emerged as a potential architecture to decouple compute and memory resources, its performance boundaries compared to standard large-scale Expert Parallelism (EP) remain underexplored. In this paper, we conduct a systematic analysis of AFD by extending the roofline model to the communication level, correlating interconnect bandwidth, arithmetic intensity, and Hardware FLOPS Utilization (HFU). Our analysis reveals a dead zone on standard clusters: increasing FFN instance count fails to improve HFU as computational workload is capped by scale-out bandwidth, causing operator active time to shrink relative to the fixed latency budget. We further show that AFD's discrete node-level scaling incurs higher imbalance penalties than EP's continuous batch adjustment. Nevertheless, these limitations diminish under specific conditions: Superpod-class hardware with abundant interconnect bandwidth and models with coarse-grained experts and lower sparsity are more likely to benefit from AFD. These findings position AFD as a promising approach for specific hardware-model combinations rather than a universal solution.

cs.DC↗

arXiv2025

ESS: An Offload-Centric Latent-Cache Management Architecture for DeepSeek-V3.2-Exp

DeepSeek-V3.2-Exp introduces a sparse attention mechanism that significantly reduces inference latency in long-context scenarios. Although the overall throughput has improved greatly, the Decode-stage of PD disaggregation remains to be a major bottleneck. This bottleneck primarily stems from the conflict between linear growth of Latent-Cache with sequence length and the limited GPU memory capacity, which constrains the feasible batch-size and thereby suppresses Decode-stage throughput. To address this challenge, we propose ESS (Extended Sparse Server), an offload-centric system design tailored for DeepSeek-V3.2-Exp. ESS selectively offloads Latent-Cache to CPU memory while preserving latency-critical components on GPU. By freeing up GPU memory, ESS effectively decoupling batch-size scaling from GPU memory constraints. This design significantly improves Decode-stage throughput, thereby reducing deployment costs in real-world settings. Our high-fidelity simulations show that ESS delivers 69.4\% throughput improvement at 32K context length and up to 123\% throughput improvement at 128K, demonstrating its effectiveness for large-context inference workloads. These results highlight ESS as a practical and scalable solution for long-context LLM serving.

cs.DC↗