Weekly Summary 2
Dec 24, 2023 12:30 pm UTC+8
科大高新区
1 本周的工作
1.1 理论调研:
- AI for system 1:
- 对整个的AI的领域分类等有个大致的分类了解, 并重构了文档,
- 回顾了LLM领域的发展脉络。
- 理清了SD的发展历程,和核心思想
- Math for compute: 数理逻辑,形式化定义。 应用数学的框架, 为了之后科研快速找到对应领域方法。
- 学术报告 Towards Efficient and Scalable Multi-GPU Computing By Xulong Tang :学术界的研究一直都是个不断向前的,就像最早专注于单核的提升,之后又偏向于多核的提升。现在对于multi-device的研究。设计的思想其实从逻辑上感觉不复杂,分析现有机器架构下应用的时间占比,关键路径,拆分对象,然后塞buffer,塞cache剪枝关键路径来加速。但是这种工作又是科研发展路上的必经之路。

- Graduation thesis 1 : 思考开题的步骤。2,3章可能要合并。
- 网站的设计通过了安老师的初步意见。
1.2 实践上手:
- SD的推理上手并简单分析热点。
1.3 其余工作:
2 下周的工作
- 逐渐深入: 是什么,运行到优化,软件优化到硬件,推理优化到训练优化,单卡到多机。
- PowerInfer 的原理,实践。
- offloader是如何决策的?判断适合PIM还是CPU。
- 他是如何理解transform的推理优化点的。
- 对比chatgpt的生成速度?
- 弄明白当下的transformer的多机推理能做到线性比吗?
- 网站的开发:
- 筛选按钮的支持,包括 news 和 events 根据type。publication根据会议筛选。
- 双语的网站支持
- 主题色的全局切换功能
- 数据的处理图片,pdf和视频存放的位置
- wiki网站,怎么公网访问
- 网站讲解的文档:
- 换admission到旧网站
- 搭建一个简化版的demo到cloudflare,event与data分离
- 面向维护者的网站逻辑,数据和文件组织逻辑
- 面向成员的文件提交逻辑。
- Math for compute: overview
- 毕业论文:标题
3 下周任务优先级
3.1 专注
- transformer -> gpt -> PowerInfer
- 论文标题
- Math for compute: overview
3.2 未知方法的工程实现
- 网站功能的开发
3.3 已知方法的工程实现
- 中文 简单版 demo github
- 文档 PPT
4 感悟
4.1 罪人舞步旋
宣布无人罪。
有种宿命感,开脱的感觉
5 Baidu ai for system
5.1 团队
- 百度编译器 晓光
- 10几个人,后端上海,前端北京。
- 尖酸小黄鸭:T9 - 半年
- T5 -清华的硕士:AI编译器前端,OP → 机器码→kernel。循环融合。
5.2 工作内容
- 鲁棒性(容忍输入扰动的稳定性,处理出错情况的可靠性),正确性,AI原生的思维。
- 后端 code阵列,
5.2.1 AI 编译器
- 中间理论:代数结构,形式化下来。多个op循环融合,map ADT(Abstract Data Type",即抽象数据类型?)
- TVM autotune op规则
- torch 负优化 - 形式化定义,算子以及约束。 A B op融合
Hi Jiang,
Happy holidays and Merry Christmas!
Here’s a summary of my activities this week:
-
AI for System 1: As a beginner in AI, I started by delving into the popular text-to-image domain. I gained insights into the fundamental concepts of stable diffusion network structures. Deploying the latest SDXL model with Comfyui on our A100 machine, I conducted image inference. Next, I plan to conduct a profiling analysis to explore inference hotspots further.
-
Graduation Thesis 1: I’ve created a basic draft of my graduation thesis outline on Overleaf and invited you to review it. I’m contemplating whether to merge chapters 2 and 3 into a single chapter.
-
ACSA Website Development 1: Professor An Hong has tentatively approved my top-down theme design, albeit with some requested changes. I anticipate spending the next few weeks meeting these requirements.
Plans for the upcoming week:
-
AI for System 2: I’ll complete a simple profile of the SDXL model’s inference. Subsequently, I’ll delve into learning transformer-like models. Once I have a foundational understanding, I plan to read the PowerInfer paper and experiment with its code. I believe the paper’s concept of offloading between GPU and CPU could be beneficial for our PIM research.
-
Graduation Thesis 2: If time permits, I aim to populate the content of the paper, starting from the background section.
-
ACSA Website Development 2: Several tasks need attention:
- Implement CN-EN language switch
- Enable global theme color change
- Tackle tasks related to adjusting font sizes for different devices.
- Determine the storage location for pictures and PDF data (consider GitHub)
- Deploy a simple demo using my top-down theme to highlight its capabilities
- Develop additional documentation and an introduction PPT to guide both website maintainers and regular users.
Looking forward to your feedback and guidance.
Best regards, Shaojie Tan