arXiv subjects

Younghun Roh

Publications and source records attributed to Younghun Roh.

3 recordsLinked to original sources

arXiv2026

Retrofitting Linear Attention into Diffusion Language Models

Diffusion language models (dLLMs) offer a promising alternative to autoregressive models by accelerating inference through parallel decoding. Recent dLLMs commonly use blockwise semi-autoregressive decoding, generating blocks autoregressively while denoising tokens within each active block in parallel. However, despite KV caching, each denoising step still attends to all previous blocks, repeatedly incurring prefix-attention cost. Motivated by this bottleneck, we ask whether dLLM inference can be further accelerated by linearizing attention over previous blocks. We introduce block-hybrid attention, which retains exact softmax attention within the active denoising block while applying linear attention over previous blocks. We show that this hybrid attention can be retrofitted into a pretrained dLLM with minimal post-training: LLaDA-Hybrid replaces 6 of the 20 attention layers in LLaDA~2.1, a 16B open-source dLLM, largely following LoLCAT (Zhang et al, 2024). The conversion takes only approximately 60 hours while preserving benchmark performance: 72.0% vs. 75.6% on HumanEval, 63.0% vs. 57.7% on MBPP+, and 86.7% vs. 88.3% on CMATH. With a Triton implementation, LLaDA-Hybrid achieves up to $1.7\times$ higher decoding throughput and supports more concurrent requests before exhausting memory, showing that pretrained dLLMs can be efficiently linearized for faster inference. Our code is available at: https://github.com/Diuven/LLaDA-Hybrid.

cs.LG↗

arXiv2026

Concurrent Balanced Augmented Trees

Augmentation makes search trees tremendously more versatile, allowing them to support efficient aggregation queries, order-statistic queries, and range queries in addition to insertion, deletion, and lookup. In this paper, we present the first lock-free augmented balanced search tree supporting generic augmentation functions. Our algorithmic ideas build upon a recent augmented unbalanced search tree presented by Fatourou and Ruppert [DISC, 2024]. We implement both data structures, solving some memory reclamation challenges in the process, and provide an experimental performance analysis of them. We also present optimized versions of our balanced tree that use delegation to achieve better scalability and performance (by more than 2x in most workloads). Our experiments show that our augmented balanced tree completes updates 2.2 to 30 times faster than the unbalanced augmented tree, and outperforms unaugmented trees by up to several orders of magnitude on 120 threads.

cs.DS↗

arXiv2025

Aggregating Funnels for Faster Fetch&Add and Queues

Many concurrent algorithms require processes to perform fetch-and-add operations on a single memory location, which can be a hot spot of contention. We present a novel algorithm called Aggregating Funnels that reduces this contention by spreading the fetch-and-add operations across multiple memory locations. It aggregates fetch-and-add operations into batches so that the batch can be performed by a single hardware fetch-and-add instruction on one location and all operations in the batch can efficiently compute their results by performing a fetch-and-add instruction on a different location. We show experimentally that this approach achieves higher throughput than previous combining techniques, such as Combining Funnels, and is substantially more scalable than applying hardware fetch-and-add instructions on a single memory location. We show that replacing the fetch-and-add instructions in the fastest state-of-the-art concurrent queue by our Aggregating Funnels eliminates a bottleneck and greatly improves the queue's overall throughput.

cs.DC↗