块稀疏注意力硬件友好却未被主流LLM采用？本文剖析原因并提出新型COBS注意力机制。
arXiv:2607.09052v1 Announce Type: new Abstract: Block sparse attention is a hardware friendly way to alleviate the key-value (KV) cache read bottlenec…
从Prefill/Decode阶段差异出发，为新兴AI加速器量身定制LLM推理性能评估方法，助力硬件选型与优化。
arXiv:2606.17104v1 Announce Type: cross Abstract: As large language models (LLMs) are increasingly deployed in latency- and cost-sensitive settings, i…
LLM推理核心瓶颈不在算力而在内存带宽，本文透彻解释为什么decode阶段受限于内存而非计算。
Article URL: https://github.com/harshuljain13/llm-inference-at-scale/blob/master/content/00_foundations/00.1_why_llm_inference_is_different/why_llm_in…
大规模GNN训练中硬件挑战的突破：Morphling通过融合不规则图遍历与密集矩阵计算实现快速灵活训练
arXiv:2512.01678v5 Announce Type: replace Abstract: Graph Neural Networks (GNNs) present a fundamental hardware challenge by fusing irregular, memory-…