전체 글 썸네일형 리스트형 [NeurIPS2025] Speculate Deep and Accurate: Lossless and Training-Free Acceleration for Offloaded LLMs via Substitute Speculative Decoding https://arxiv.org/abs/2509.18344, NeurIPS 2025 Published Paper Abstract메모리가 제한된 단일 GPU 환경에서 LLM을 추론하기 위해서 주로 다음 두 가지 접근법이 활용됨.1) 모델 양자화(Quantization): 모델의 메모리 풋프린트를 줄일 수 있으나, 양자화 오차(Quantization Error)로 인한 생성 품질 저하가 발생할 수 있음. 2) 파라미터 오프로딩(Parameter Offloading): 품질 저하는 방지하지만, CPU-GPU 간 데이터 전송 병목으로 인해 추론 속도가 크게 저하됨. Speculative Decoding (SD)은 오프로딩 환경에서 추론을 가속하기 위한 기법으로, 파라미터 수가 적은 Draft 모델과 원본 Tar.. 더보기 [ICLR 2026] When Reasoning Meets Compression: Understanding the Effects of LLMs Compression on Large Reasoning Models https://arxiv.org/abs/2504.02010, ICLR 2026 When Reasoning Meets Compression: Understanding the Effects of LLMs Compression on Large Reasoning ModelsCompression methods, including quantization, distillation, and pruning, improve the computational efficiency of large reasoning models (LRMs). However, existing studies either fail to sufficiently compare all three compression methods on LRMs or lac.. 더보기 [ICML 2025] Q-VDiT: Towards Accurate Quantization and Distillation of Video-Generation Diffusion Transformers https://arxiv.org/abs/2505.22167 ICML 2025 Accepted Paper Q-VDiT: Towards Accurate Quantization and Distillation of Video-Generation Diffusion TransformersDiffusion transformers (DiT) have demonstrated exceptional performance in video generation. However, their large number of parameters and high computational complexity limit their deployment on edge devices. Quantization can reduce storage r.. 더보기 [ICLR 2025] DGQ: Distribution-Aware Group Quantization for Text-to-Image Diffusion Models https://arxiv.org/abs/2501.04304 AbstractText-to-Image 생성을 위한 Diffusion모델은 높은 계산량과 메모리사용으로 인해 실제 적용에 제약을 줌.이를 하기 위해 Quantization 기법이 이용되는데, 기존 Diffusion모델의 quantization은 낮은 비트에서 이미지 품질과 텍스트-이미지 alignment를 유지하는데 한계를 지님.본 논문에서는 activation에서 outlier가 있으며, 이는 이미지 품질을 결정하는데 중요한 역할을 한다는 것을 분석함. 또한, 텍스트-이미지 alignment에 cross-attention이 중요한 역할을 한다는 것을 분석함.논문에서 제안하는 DGQ(Distribution-aware Group Quantizat.. 더보기 [ICLR 2025] ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation https://arxiv.org/abs/2406.02540 ICLR 2025 ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video GenerationDiffusion transformers have demonstrated remarkable performance in visual generation tasks, such as generating realistic images or videos based on textual instructions. However, larger model sizes and multi-frame processing for video generation lead to i.. 더보기 [CVPR 2025] Q-DiT: Accurate Post-Training Quantization for Diffusion Transformers https://arxiv.org/abs/2406.17343 , CVPR 2025 (Accepted) Q-DiT: Accurate Post-Training Quantization for Diffusion TransformersRecent advancements in diffusion models, particularly the architectural transformation from UNet-based models to Diffusion Transformers (DiTs), significantly improve the quality and scalability of image and video generation. However, despite their impressiarxiv.org Abstrac.. 더보기 [ECCV2024] Post-training Quantization for Text-to-Image Diffusion Models with Progressive Calibration and Activation Relaxing https://arxiv.org/abs/2311.06322 , ECCV 2024 Post-training Quantization for Text-to-Image Diffusion Models with Progressive Calibration and Activation RelaxingHigh computational overhead is a troublesome problem for diffusion models. Recent studies have leveraged post-training quantization (PTQ) to compress diffusion models. However, most of them only focus on unconditional models, leaving the q.. 더보기 [ACM SAC 2025] Advanced Knowledge Transfer: Refined Feature Distillation for Zero-Shot Quantization in Edge Computing https://dl.acm.org/doi/abs/10.1145/3672608.3707816Abstract기존 Zero-Shot Quantization(ZSQ, Data-Free Quantizaton) 분야에서는 full-precision(FP) Model로부터 높은 quality의 데이터를 생성하는 데 초점을 두는 연구가 진행되고 있음. 하지만, low-bit(높은 압축률) 환경에서 Quantized Model을 학습할 때는 Quantized Model이 정보 수용량 관련 한계를 갖기 때문에 데이터를 생성하는 기법만으로는 적절한 학습이 이루어지지 않음. 이러한 한계를 개선하기 위해 본 논문에서는 Quantized Model을 효과적으로 학습하기 위한 AKT(Advanced Knowledge Transfe.. 더보기 이전 1 2 3 4 ··· 7 다음
