Xuming Ye | Searcharxiv

arXiv2026

GRAFT: Adaptive DLM-Based Draft Tree Construction with Target-Distilled Edge Scoring

Tree-based speculative decoding raises the mean accepted tokens of standard speculative decoding by verifying multiple draft paths, and existing tree builders typically construct these paths through parent-conditioned expansion, where each child token is generated conditioned on its parent path. This construction is incompatible with diffusion language model (DLM) drafters such as DFlash, which produces all future-position distributions in a single forward pass. DDTree bridges this gap by treating high-probability tokens from each future-position distribution as candidate nodes and selecting edges between consecutive positions under a fixed node budget. However, its edge selection relies on token probability alone without modeling parent--child compatibility, so target-compatible tokens can be attached to wrong parents; moreover, its fixed budget ignores that the throughput-optimal tree size varies with the decoding state. We propose GRAFT, a draft-tree construction framework for DLM-based speculative decoding. GRAFT introduces Target-Distilled Edge Scoring (TDES), which distills parent--child preferences from target-model traces to select target-compatible edges, and State-Aware Budget Allocation (SABA), which sets the per-round tree budget by balancing expected draft gain against verification cost. Across multiple models and tasks, GRAFT achieves $2.13\times$--$6.36\times$ end-to-end speedup over autoregressive decoding while adding less than $0.5$\,ms of overhead per round, approximately $1.4\%$ of the target-model verification latency.

cs.CL↗

arXiv2021

Chunk Content is not Enough: Chunk-Context Aware Resemblance Detection for Deduplication Delta Compression

With the growing popularity of cloud storage, removing duplicated data across users is getting more critical for service providers to reduce costs. Recently, Data resemblance detection is a novel technology to detect redundancy among similarity. It extracts feature from each chunk content and treat chunks with high similarity as candidates for removing redundancy. However, popular resemblance methods such as "N-transform" and "Finesse" use only the chunk data for feature extraction. A minor modification on the data chunk could seriously deteriorate its capability for resemblance detection. In this paper, we proposes a novel chunk-context aware resemblance detection algorithm, called CARD, to mitigate this issue. CARD introduces a BP-Neural network-based chunk-context aware model, and uses N-sub-chunk shingles-based initial feature extraction strategy. It effectively integrates each data chunk content's internal structure with the context information for feature extraction, the impact of small changes in data chunks is significantly reduced. To evaluate its performance, we implement a CARD prototype and conduct extensive experiments using real-world data sets. The results show that CARD can detect up to 75.03% more redundant data and accelerate the resemblance detection operations by 5.6 to 17.8 times faster compared with the state-of-the-art resemblance detection approaches.

cs.DC↗