arXiv subjects

Kangyuan Shu

Publications and source records attributed to Kangyuan Shu.

3 recordsLinked to original sources

arXiv2026

Self-Speculative Biased Decoding for Faster Re-Translation

Large language models achieve strong machine translation quality but incur high inference cost and latency, posing challenges for simultaneous translation. Re-translation provides a practical solution for off-the-shelf LLMs by repeatedly regenerating the target output as the source input grows, but it suffers from substantial redundant computation. We propose Self-Speculative Biased Decoding (SSBD), a simple and tuning-free inference method that accelerates re-translation by exploiting temporal coherence in streaming translation. SSBD reuses the model's previous output as a speculative draft for the updated input, verifies the draft efficiently in a single forward pass with a lightweight bias, and resumes autoregressive decoding only from the first divergence. We further introduce a display-only masking strategy that hides unstable suffixes from the user interface while retaining them in the draft for verification and potential acceptance. Experiments show that SSBD achieves substantial speedup over standard re-translation while maintaining comparable translation quality, without architectural changes, auxiliary models, or extra fine-tuning.

cs.CL↗

arXiv2025

Enabling Dynamic Sparsity in Quantized LLM Inference

Deploying large language models (LLMs) on end-user devices is gaining importance due to benefits in responsiveness, privacy, and operational cost. Yet the limited memory and compute capability of mobile and desktop GPUs make efficient execution difficult. Recent observations suggest that the internal activations of LLMs are often dynamically sparse, meaning that for each input, only part of the network contributes significantly to the output. Such sparsity could reduce computation, but it interacts poorly with group-wise quantization, which remains the dominant approach for fitting LLMs onto resource-constrained hardware. To reconcile these two properties, this study proposes a set of techniques that realize dynamic sparse inference under low-bit quantization. The method features: (1) a zigzag-patterned quantization layout that organizes weights in a way consistent with activation sparsity and improves GPU memory locality; (2) a specialized GEMV kernel designed for this layout to fully utilize parallel compute units; and (3) a compact runtime mechanism that gathers sparse indices with minimal overhead. Across several model scales and hardware configurations, the approach achieves up to 1.55x faster decoding throughput while maintaining accuracy comparable to dense quantized inference, showing that structured sparsity and quantization can effectively coexist on commodity GPUs.

cs.DC↗

arXiv2022

Streaming, fast and accurate on-device Inverse Text Normalization for Automatic Speech Recognition

Automatic Speech Recognition (ASR) systems typically yield output in lexical form. However, humans prefer a written form output. To bridge this gap, ASR systems usually employ Inverse Text Normalization (ITN). In previous works, Weighted Finite State Transducers (WFST) have been employed to do ITN. WFSTs are nicely suited to this task but their size and run-time costs can make deployment on embedded applications challenging. In this paper, we describe the development of an on-device ITN system that is streaming, lightweight & accurate. At the core of our system is a streaming transformer tagger, that tags lexical tokens from ASR. The tag informs which ITN category might be applied, if at all. Following that, we apply an ITN-category-specific WFST, only on the tagged text, to reliably perform the ITN conversion. We show that the proposed ITN solution performs equivalent to strong baselines, while being significantly smaller in size and retaining customization capabilities.

cs.CL↗