Jae‐Han Park

Hyucksung Kwon, Kyungmo Koo, Janghyeon Kim et al.

The expansion of large language models (LLMs) with hundreds of billions of parameters presents significant challenges to computational resources, particularly data movement and memory bandwidth. Long-context LLMs, which process sequences of tens of thousands of tokens, further increase the demand on the memory system as the complexity in attention layers and key-value cache sizes is proportional to the context length. Processing-in-Memory (PIM) maximizes memory bandwidth by moving compute to the data and can address the memory bandwidth challenges; however, PIM is not necessarily scalable to accelerate long-context LLM because of limited per-module memory capacity and the inflexibility of fixed-functional unit PIM architecture and static memory management. In this work, we propose LoL-PIM which is a multi-node PIM architecture that accelerates long context LLM through hardware-software co-design. In particular, we propose how pipeline parallelism can be exploited across a multi-PIM module while a direct PIM access (DPA) controller (or DMA for PIM) is proposed that enables dynamic PIM memory management and results in efficient PIM utilization across a diverse range of context length. We developed an MLIR-based compiler for LoL-PIM extending a commercial PIM-based compiler where the software modifications were implemented and evaluated, while the hardware changes were modeled in the simulator. Our evaluations demonstrate that LoL-PIM significantly improves throughput and reduces latency for long-context LLM inference, outperforming both multi-GPU and GPU-PIM systems (up to 8.54x and 16.0x speedup, respectively), thereby enabling more efficient deployment of LLMs in real-world applications.

Ha-Yeong Choi, Jaehan Park

Despite remarkable advancements in recent voice conversion (VC) systems, enhancing speaker similarity in zero-shot scenarios remains challenging. This challenge arises from the difficulty of generalizing and adapting speaker characteristics in speech within zero-shot environments, which is further complicated by mismatch between the training and inference processes. To address these challenges, we propose VoicePrompter, a robust zero-shot VC model that leverages in-context learning with voice prompts. VoicePrompter is composed of (1) a factorization method that disentangles speech components and (2) a DiT-based conditional flow matching (CFM) decoder that conditions on these factorized features and voice prompts. Additionally, (3) latent mixup is used to enhance in-context learning by combining various speaker features. This approach improves speaker similarity and naturalness in zero-shot VC by applying mixup to latent representations. Experimental results demonstrate that VoicePrompter outperforms existing zero-shot VC systems in terms of speaker similarity, speech intelligibility, and audio quality. Our demo is available at \url{https://hayeong0.github.io/VoicePrompter-demo/}.

Yeong-Jun Cho, Su-A Kim, Jae-Han Park et al.

Person re-identification is the task of recognizing or identifying a person across multiple views in multi-camera networks. Although there has been much progress in person re-identification, person re-identification in large-scale multi-camera networks still remains a challenging task because of the large spatio-temporal uncertainty and high complexity due to a large number of cameras and people. To handle these difficulties, additional information such as camera network topology should be provided, which is also difficult to automatically estimate, unfortunately. In this study, we propose a unified framework which jointly solves both person re-identification and camera network topology inference problems with minimal prior knowledge about the environments. The proposed framework takes general multi-camera network environments into account and can be applied to online person re-identification in large-scale multi-camera networks. In addition, to effectively show the superiority of the proposed framework, we provide a new person re-identification dataset with full annotations, named SLP, captured in the multi-camera network consisting of nine non-overlapping cameras. Experimental results using our person re-identification and public datasets show that the proposed methods are promising for both person re-identification and camera topology inference tasks.

Yeong-Jun Cho, Jae-Han Park, Su-A Kim et al.

Person re-identification in large-scale multi-camera networks is a challenging task because of the spatio-temporal uncertainty and high complexity due to large numbers of cameras and people. To handle these difficulties, additional information such as camera network topology should be provided, which is also difficult to automatically estimate. In this paper, we propose a unified framework which jointly solves both person re-id and camera network topology inference problems. The proposed framework takes general multi-camera network environments into account. To effectively show the superiority of the proposed framework, we also provide a new person re-id dataset with full annotations, named SLP, captured in the synchronized multi-camera network. Experimental results show that the proposed methods are promising for both person re-id and camera topology inference tasks.