arXiv subjects

Bingyao Li

Publications and source records attributed to Bingyao Li.

4 recordsLinked to original sources

arXiv2026

Line-Coupled Language Model

Autoregressive language models generate one token per decoding step, limiting the useful output of each forward pass. Although diffusion models, insertion-based decoding, and multi-token prediction enable parallel generation, they either incur additional training-time token traffic or struggle to predict strongly dependent future tokens. We introduce the Line-Coupled Language Model (LCLM), an autoregressive model that advances multiple text lines together by predicting the next token for every active line while coupling the lines through shared causal context. LCLM interleaves line tokens into a single causal sequence and uses line-staggered rotary positions, retaining the standard next-token objective and causal attention. Controlled experiments show that cross-line targets are substantially less dependent than consecutive same-line targets, supporting lines as parallel generation units. With 881M parameters, LCLM produces an average of 2.94 content tokens per forward pass with a validation cross-entropy loss of 2.44, compared with 1.00 token per forward pass and a loss of 2.39 for the vanilla autoregressive baseline. Most notably, even when LCLM generates 16 tokens per forward pass, its loss is only 0.09 higher than that of the vanilla autoregressive baseline (2.34 vs. 2.25).

cs.CL↗

arXiv2026

SkyVLaM: Multimodal Large Language Model for UAV Video Understanding in Remote Sensing

Recent advances in Multimodal Large Language Models (MLLMs) have significantly improved remote sensing (RS) multimodal understanding. Language-conditioned segmentation is crucial for fine-grained target understanding in Unmanned Aerial Vehicle (UAV) videos. However, this task remains challenging due to the prevalence of small, visually ambiguous targets and dynamic aerial perspectives. In this paper, we propose SkyVLaM, a multimodal large language model for UAV video understanding. SkyVLaM constructs sparse tokens directly from patch-level video representations through a temporal basis perceiver, regularizes the sparse basis to encourage complementary temporal cues, and adaptively selects a temporally coherent dense segment for high-resolution inspection. The resulting sparse and dense tokens are jointly processed by a large language model for query-conditioned segmentation. We further build SkyVid, consisting of SkyVid-VGCG and SkyVid-RVOS for video grounded conversation generation and referring video object segmentation, respectively. SkyVid contains 101 videos, 33.6K frames, and 1.53M pixel-level object instances. Experiments show that SkyVLaM provides a more effective allocation of the visual token budget and improves language-conditioned video segmentation in UAV scenarios.

cs.CV↗

arXiv2024

Improving GPU Multi-Tenancy Through Dynamic Multi-Instance GPU Reconfiguration

Continuous learning (CL) has emerged as one of the most popular deep learning paradigms deployed in modern cloud GPUs. Specifically, CL has the capability to continuously update the model parameters (through model retraining) and use the updated model (if available) to serve overtime arriving inference requests. It is generally beneficial to co-locate the retraining and inference together to enable timely model updates and avoid model transfer overheads. This brings the need for GPU sharing among retraining and inferences. Meanwhile, multiple CL workloads can share the modern GPUs in the cloud, leading to multi-tenancy execution. In this paper, we observe that prior GPU-sharing techniques are not optimized for multi-tenancy CL workloads. Specifically, they do not coherently consider the accuracy of the retraining model and the inference service level objective (SLO) attainment. Moreover, they cannot accommodate the overtime dynamics (e.g., inference arrival intensity) in CL execution. In this paper, we propose MIGRator, a novel GPU reconfiguration runtime that dynamically performs GPU reconfiguration for multi-tenancy CL workloads. MIGRator is based on the recent NVIDIA multi-instance GPU (MIG) to mitigate resource contention and formulates the reconfiguration optimization into Integer Linear Programming (ILP) to dynamically identify, reconfigure, and allocate the GPU instances. MIGRator leverages the "Goodput" metric in the ILP objective function to consider both inference SLO attainment and model accuracy in the reconfiguration exploration. We evaluate MIGRator using representative multi-tenancy CL workloads. The results show our approach outperforms the state-of-the-art GPU sharing techniques (i.e., Ekya, Astraea, and PARIS) by 17\%, 21\%, and 20\%, respectively.

cs.DC↗

arXiv2024

Improving Multi-Instance GPU Efficiency via Sub-Entry Sharing TLB Design

NVIDIA's Multi-Instance GPU (MIG) technology enables partitioning GPU computing power and memory into separate hardware instances, providing complete isolation including compute resources, caches, and memory. However, prior work identifies that MIG does not extend to partitioning the last-level TLB (i.e., L3 TLB), which remains shared among all instances. To enhance TLB reach, NVIDIA GPUs reorganized the TLB structure with 16 sub-entries in each L3 TLB entry that have a one-to-one mapping to the address translations for 16 pages of size 64KB located within the same 1MB aligned range. Our comprehensive investigation of address translation efficiency in MIG identifies two main issues caused by L3 TLB sharing interference: (i) it results in performance degradation for co-running applications, and (ii) TLB sub-entries are not fully utilized before eviction. Based on this observation, we propose STAR to improve the utilization of TLB sub-entries through dynamic sharing of TLB entries across multiple base addresses. STAR evaluates TLB entries based on their sub-entry utilization to optimize address translation storage, dynamically adjusting between a shared and non-shared status to cater to current demand. We show that STAR improves overall performance by an average of 30.2% across various multi-tenant workloads.

cs.DC↗