Mingzhe Huang

Mingzhe Huang, Weijun Wang, Xin Ding et al.

In Vision-Language Models (VLMs), processing a massive number of visual tokens incurs prohibitive computational overhead. While recent training-aware pruning methods attempt to selectively discard redundant tokens, they largely rely on continuous-gradient relaxations. However, visual token pruning is inherently a discrete, non-convex combinatorial problem; consequently, these continuous approximations frequently trap the optimization in sub-optimal local minima, especially under aggressive compression budgets. To overcome this fundamental bottleneck, we propose GRIP-VLM, a Group-Relative Importance Pruning framework driven by Reinforcement Learning. Rather than relying on smooth-gradient assumptions, GRIP-VLM formulates pruning as a Markov Decision Process, employing a Group Relative Policy Optimization (GRPO) paradigm anchored by supervised warm-up to directly explore the discrete selection space. Integrated with a budget-aware scorer, our lightweight agent dynamically evaluates per-token importance and adapts to arbitrary compression ratios without retraining. Extensive experiments across diverse multimodal benchmarks demonstrate that GRIP-VLM consistently outperforms heuristic and supervised-learning baselines, achieving a superior Pareto frontier and delivering up to a 15\% inference speedup at equal accuracy.

Yuchen Luo, Fangyue Zhu, Ruining Zhou et al.

Post-Training Quantization (PTQ) is crucial for efficient model deployment, yet its effectiveness on Ascend NPU remains under-explored compared to GPU architectures. This paper presents a case study of representative PTQ baselines applied to reasoning-oriented models such as DeepSeek-R1-Distill-Qwen series (1.5B/7B/14B) and QwQ-32B. We evaluate four distinct algorithms, including AWQ, GPTQ, SmoothQuant, and FlatQuant, to cover the spectrum from weight-only compression to advanced rotation-based methods. Our empirical results reveal significant platform sensitivity. While 4-bit weight-only quantization proves viable for larger models, aggressive 4-bit weight-activation schemes suffer from layer-wise calibration instability on the NPU, leading to logic collapse in long-context reasoning tasks. Conversely, standard 8-bit quantization remains numerically stable. Furthermore, a real-world INT8 deployment demonstrates that although optimized kernels reduce latency, dynamic quantization overheads currently limit end-to-end acceleration. These findings offer a practical reference for the feasibility and limitations of deploying quantized reasoning models on Ascend NPU.

Huan Yang, Renji Zhang, Mingzhe Huang et al.

Recent advances in long-text understanding have pushed the context length of large language models (LLMs) up to one million tokens. It boosts LLMs's accuracy and reasoning capacity but causes exorbitant computational costs and unsatisfactory Time to First Token (TTFT). KV cache reuse, which reuses the exact same KV cache of prefixes and templates or shares similar ones but with extra selective recomputation, offers a promising way to tackle this issue. However, prior studies overlook the cross-request KV reuse and the attention deviations introduced by new tokens during the decoding stage. In this paper, we present a KV cache management module that shares the KV cache across requests under multi-tenant scenarios without sacrificing model accuracy. Our system, KVShare, enables accurate and efficient LLM serving by 1) a Dual-Stage High Deviation algorithm (DHD) that conditionally selects a small portion of KV cache to be recomputed during both prefill and decode phases, and 2) a cache-aware scheduler that prioritizes requests based on their KV cache hit rates and orchestrates continuous batching to achieve enhanced system efficiency and faster TTFT. Multi-task experiments conducted on models such as Qwen2.5-7B,Llama3.1-8B and Yi1.5-9B demonstrate that KVShare reduces TTFT by up to 9.39x and increases 1.2x of the throughput compared to the full KV recompute. Moreover, KVShare achieves 20.38% boost in terms of accuracy compared to SOTA methods.