Tag
This paper systematically studies variants of QKV projection sharing in transformers, finding that sharing key and value projections (Q-K=V) achieves 50% KV cache reduction with only 3.1% perplexity degradation, and combining with GQA/MQA can reach up to 96.9% cache reduction—enabling practical on-device inference with minimal quality loss.
Complexity-Balanced Splitting (CBS) partitions the diffusion timeline into segments of equal approximation burden using local complexity measures, improving synthesis quality by ~35% in FID without increasing inference cost.
A new method called Zero-Expert Self-Distillation Adaptation (ZEDA) allows MoE models like Qwen3 and GLM to skip half their expert computations on easy tokens with minimal accuracy loss, achieving ~20% inference speedup by adding dummy experts that output nothing.
Q-ARVD is a novel quantization framework to reduce inference costs of autoregressive video diffusion models by addressing frame-wise sensitivity imbalance and weight outlier patterns.
ZEDA is a low-cost framework that converts post-trained static MoE models into dynamic ones by injecting zero-output experts and using self-distillation, achieving over 50% expert FLOP reduction with marginal accuracy loss on benchmarks.
This paper investigates the parameter-level mechanisms behind the efficiency of On-Policy Distillation (OPD) for large language models, attributing it to early 'foresight' in module allocation and update direction. It proposes EffOPD, a plug-and-play method that accelerates OPD training by 3x without compromising final performance.
SlimSpec introduces a low-rank parameterization for drafter LM-heads to accelerate speculative decoding in LLMs, achieving 4-5x speedup while maintaining full vocabulary support.
This research paper analyzes the internal mechanics of Large Vision-Language Models (LVLMs) using information theory, revealing that attention mechanisms may be redundant while Feed-Forward Networks drive semantic innovation. The authors demonstrate that replacing learned attention weights with random values can yield comparable performance, suggesting current models 'get lost in attention'.