9 papers
Compass: Dissecting Communication and Computation Operators for Efficient LLM Training
Guangyu Xiang, Lin Zhang, Haoxuan Yu +3
Overlapping communication and computation operators is a common practice to hide communication overheads, accelerating large language models (LLMs) training on GPU clusters. Existi…
SAKI-RoPE: Score-Aware Low-Rank Key Indexing with RoPE-Aware Covariance and Random-Matrix Noise Analysis
Lin Zhang
Existing low-rank KV-cache indexes preserve model weights or key variance, not the attention scores that determine retrieval. We derive the expected score distortion of key-side ra…
Zellige: Moldable Sequence Placement for Mixed Image-Video DiT Training
Guangyu Xiang, Xueze Kang, Minwei Zhao +4
High-quality video generation requires training Diffusion Transformers (DiTs) jointly on image and video data, posing a mixed-length sequence training problem across GPUs. Existing…
Xema: Efficient Diffusion Serving through Fine-Grained Memory Management and Auto-Configuration
Xueze Kang, Guangyu Xiang, Suyi Li +4
Diffusion models are increasingly deployed as production visual-generation services, where serving high-resolution image and long video generation is often limited by GPU memory. P…
KernelFlume: Elastic Core-Attention Scaling for Agentic Long-Context Decoding
Guangyu Xiang, Xueze Kang, Lin Zhang +4
LLM serving is increasingly dominated by long and dynamic decode workloads from agents, reasoning models, and extended conversations. When bursty long-context demand exceeds deploy…
HyLRA: Hybrid Layer Reuse Attention for Efficient Long-Context Inference
Xuan Ai, Qingqing Yang, Peng Wang +4
Long-context inference in Large Language Models (LLMs) is bottlenecked by the quadratic computation complexity of attention and the substantial memory footprint of Key-Value (KV) c…
HierMoE: Accelerating MoE Training with Hierarchical Token Deduplication and Expert Swap
Wenxiang Lin, Xinglin Pan, Lin Zhang +3
The sparsely activated mixture-of-experts (MoE) transformer has become a common architecture for large language models (LLMs) due to its sparsity, which requires fewer computationa…
BigMac: A Communication-Efficient Mixture-of-Experts Model Structure for Fast Training and Inference
Zewen Jin, Shengnan Wang, Jiaan Zhu +5
The Mixture-of-Experts (MoE) structure scales the Transformer-based large language models (LLMs) and improves their performance with only the sub-linear increase in computation res…
XL3M: A Training-free Framework for LLM Length Extension Based on Segment-wise Inference
Shengnan Wang, Youhui Bai, Lin Zhang +7
Length generalization failure problem, namely the large language model (LLM) fails to generalize to texts longer than its maximum training length, greatly restricts the application…