MINEDRAFT: A Framework for Batch Parallel Speculative Decoding 把推测解码打成mini batch,随后在drafter和verifier上分批处理。 在vllm上修改,工程量可观。不错的尝试和idea。 Architecture overview of MINEDRAFT. (Left) The Scheduler manages r
未来的硬件怎么应对不断演变的软件? https://dl.acm.org/doi/pdf/10.1145/3695053.3731014 这篇文章是 ISCA 2025 的论文 《Neoscope: How Resilient Is My SoC to Workload Churn?》,核心在回答一个非常系统/架构导向的问题: 当软件和工作负载不断演进(churn)时,一个 SoC 设计在整个生命
今天yf来分享一篇来自IPADS的ATC25文章。 Colocating ML Inference and Training with Fast GPU Memory Handover 简短点评:依旧IPADS特有的大工程,TVM+vLLM+NCCL+Pytorch 开组会大家一起问了很多问题。 https://ipads.se.sjtu.edu.cn/_media/publications/si
STOC81 I/O Complexity: The Red-Blue Pebble Game 这是一篇理论计算机科学文章,但是描述了一个非常有趣的问题:就像时间复杂度一样,我们能不能做一个I/O复杂度,衡量一个程序最少要进行多少次I/O? 文章链接: https://www.eecs.harvard.edu/~htk/publication/1981-stoc-hong-kung.pdf Com
Feedback Feedback to Learner 12/30/25 3:55 PM 82+5=87 (extra: 0) > Summary: As we demonstrated in the lab, you should pre-assign labels and taints to cluster nodes using Kind config YAML. Other parts
This is the self-review pack of Distributed Systems and Cloud Computing. We have lesson 1-5. Lesson 1 Presentation – Effective communication of information rather than of data – Code and number conver
In this assignment, we need to use Spark to analyze the Parking dataset. Preparing Install pysark and java pip install pyspark sudo apt-get update sudo apt-get install openjdk-17-jdk export JAVA_HOME=
Scores 95+10=105 (extra: 5) Summary: The impl is nice in general, and the report is awesome! Yes, this is an assignment where you should follow certain instructions and submit certain stuff, but just
你和你的研究 https://gwern.net/doc/science/1986-hamming Great work is something else than mere brains. Brains are measured in various ways. In mathematics, theoretical physics, astrophysics, typically brain
Can We Trust Profiling Results? Understanding and Fixing the Inaccuracy in Modern Profilers https://par.nsf.gov/servlets/purl/10122098 在上次阅读完博客 # Where Do Interrupts Happen? 后(我的中文解析:https://www.haibi
This paper is in arxiv for 2 years. Then it goes into NSDI26. Maybe we can see the difference between versions of 2023 and 2026. Paper link: https://arxiv.org/pdf/2305.05920 Main idea: Can we use MLFQ