2026 Spring
Specific Requirements
-
We adopt a topic-based organization this semester. Presenters should select papers that fit the semester topics. Paper sources are no longer restricted to SOSP, OSDI, or any specific venues, as long as the paper is confirmed with the organizers at registration time and meets the quality bar.
-
The presentation format is flexible. For one paper, a full discussion should last 45โ50 minutes, while a sharing presentation should last 30โ35 minutes. Presenters are expected to prepare around 30โ40 slides.
-
The presentation should clearly introduce the main idea, technical contributions, and key strengths and weaknesses of the paper. Relevant background and related work may also be included when appropriate.
Other Information
The playback video and text summary will be uploaded to bilibili and zhihu as soon as possible.
Schedule
April 7
- ๐ก Kick-off meeting
- ๐โโ๏ธ Zewen Jin, Chizheng Fang, Yuzhe Li, Mulong Li and Shen Fu
- ๐ slides
April 14
Topic I
- ๐ก [CVPR'26] AdaCluster: Adaptive Query-Key Clustering for Sparse Attention in Video Generation
- ๐โโ๏ธ Haoyue Tan
- ๐ slides
- ๐ Q&A summary, ๐บ video
Topic II
- ๐ก [arXiv] IndexCache: Accelerating Sparse Attention via Cross-Layer Index Reuse
- ๐โโ๏ธ Ruibo Liu, Ouxiang Zhou
- ๐ slides
- ๐บ video
April 21
Topic I
- ๐ก [arXiv] Token Sparse Attention: Efficient Long-Context Inference with Interleaved Token Selection
- ๐โโ๏ธ Chengjie Tang, Shen Fu
- ๐ slides
- ๐บ video
April 28
Topic I
- ๐ก [arXiv] PROBE: Co-Balancing Computation and Communication in MoE Inference via Real-Time Predictive Prefetching
- ๐โโ๏ธ Qinghe Wang
- ๐ slides
Topic II
- ๐ก [arXiv] FluxMoE: Decoupling Expert Residency for High-Performance MoE Serving
- ๐โโ๏ธ Long Zhao
- ๐ slides
May 12
- ๐ก Comprehensive introduction of DeepSeek-V4 technical report (PART I)
- ๐โโ๏ธ Chengru Yang, Chengjie Tang, Ouxiang Zhou, Shen Fu, Ruibo Liu, Yinhe Chen
- ๐ slides
May 19
- ๐ก Comprehensive introduction of DeepSeek-V4 technical report (PART II)
- ๐โโ๏ธ Congkun Ai, Yuzhe Li, Jiahui Tan, Chenhan Wang, Chizheng Fang
- ๐ slides
May 26
- ๐ก [arXiv] JANUS: Disaggregating Attention and Experts for Scalable MoE Inference
- ๐โโ๏ธ Chizheng Fang
- ๐ slides
June 2
Topic I
- ๐ก [MLSys26] AccelOpt: Self-improving Agents for AI Accelerator Kernel Optimization
- ๐โโ๏ธ Yuhang Wang, Jingwen Sun
- ๐ slides
Topic II
- ๐ก [arXiv] CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generation
- ๐โโ๏ธ Xingye You, Ziqi Chen
June 9
Topic I
- ๐ก [ICLR26] SonicMoE: Accelerating MoE with IO and Tile-aware Optimizations
- ๐โโ๏ธ Jiahui Tan
Topic II
- ๐ก [PLDI26] Neptune: Advanced ML Operator Fusion for Locality and Parallelism on GPUs
- ๐โโ๏ธ Congkun Ai, Zewen Jin
June 16
Topic I
- ๐ก ๅบไบๆ่ พๅ็ฆปๅผๆถๆ็Attention่ๅ็ฎๅญไผๅ
- ๐โโ๏ธ Huaman ZHou
Topic II
- ๐ก [MLSys26] BLASST: Dynamic BLocked Attention Sparsity via Softmax Thresholding
- ๐โโ๏ธ Zijian Dai, Sen Han
- ๐ slides
June 23
Topic I
- ๐ก [SIGMOD26] Twenty Years of Bigtable
- ๐โโ๏ธ Chuannan Zhang
- ๐ slides
June 30
Topic I
- ๐ก [FAST26] OdinANN: Direct Insert for Consistently Stable Performance in Billion-Scale Graph-Based Vector Search
- ๐โโ๏ธ Mulong Li
- ๐ slides
Topic II
- ๐ก [SOSP25] Aegaeon: Effective GPU Pooling for Concurrent LLM Serving on the Market
- ๐โโ๏ธ Bosen Yang, Jiyang Wang
- ๐ slides
July 7
Topic I
- ๐ก [arXiv] ReMP: Low-Downtime Runtime Model-Parallelism Reconfiguration for LLM Serving
- ๐โโ๏ธ Jiaan Zhu
- ๐ slides
Topic II
- ๐ก [OSDI25] To PRI or Not To PRI, Thatโs the question
- ๐โโ๏ธ Yicheng Zhang, Luofan Chen
- ๐ slides
July 14
Topic I
- ๐ก [OSDI26] Strata: Hierarchical Context Caching for Long Context Language Model Serving
- ๐โโ๏ธ Mingxuan Liu
- ๐ slides
Topic II
- ๐ก [MLSys26] fabric-lib: RDMA Point-to-Point Communication for LLM Systems
- ๐โโ๏ธ Mingxuan Liu
- ๐ slides