Initial commit: obsidian to gitea

This commit is contained in:
2026-05-07 15:04:41 +08:00
commit a57afa86b4
323 changed files with 42569 additions and 0 deletions

Binary file not shown.

After

Width:  |  Height:  |  Size: 2.1 MiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 1.2 MiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 2.2 MiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 2.0 MiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 2.0 MiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 2.1 MiB

View File

@@ -0,0 +1,58 @@
- [x] HybriMoE: Hybrid CPU-GPU Scheduling and Cache Management for Efficient MoE Inference
- [ ] MILLION: MasterIng Long-Context LLM Inference Via Outlier-Immunized KV Product QuaNtization
- [ ] Multiplexing Dynamic Deep Learning Workloads with SLO-awareness in GPU Clusters
- [x] SimAI: 面向 AI 大规模集群的高精度仿真器
scale up: the bigger, the better
- [ ] SoMa:深度神经网络加速器 DRAM 通信调度空间的识别、探索与理解
---
内存存储优化:SSD 结构优化与落地的问题
- 面向多核系统的高性能高可靠异步通信
![[250524-113000.jpeg]]
衡量 AI 赋能开发的角度:AI 生成代码的上库率
mini panel
- startup 短板要长,否则会被 judge,市面上的产品功能,你做的不好
- 形式化验证的速度慢,与开发速度之间的 gap
KTransformers
- AMX 指令加速 CPU 上的计算
---
吞吐 = 时延 * 处理速度
世界本质是稀疏的:MoE
![[250525-095739.jpeg]]
训练时 lookahead,使得支持 MTP
![[250525-100200.jpeg]]
逐请求分布式(DP) -> 逐阶段分离(PD 分离) -> 逐层分布式(DeepEP、注意力卸载)
![[250525-100839.jpeg]]
网络:全双工;内存读写:半双工
**当前网络带宽已经超过内存带宽**
目标:EP 通信量均衡和 DP 计算量均衡
![[250525-101147.jpeg]]
![[250525-101629.jpeg]]
2000+ H800,服务国内外全部 DeepSeek 流量,春节过后做了吞吐优化牺牲了延迟优化
EP 动态扩缩容,EP32 -> EP320,PD role 的动态转化

34
study/conf/EuroSys'26.md Normal file
View File

@@ -0,0 +1,34 @@
### Plan
#### 4.27: Workshops
Kilsyth Level 0
EuroMLSys
#### 4.28
13:30 – 15:15
Track A — LLM Training at Scale
19:15 – 22:15   Banquet (NMS)
#### 4.29
8:45 – 10:15
AI
13:30 – 15:15
==Track A — LLM Inference==
- High Throughput and Low Latency LLM Serving via Adaptive KV Caching
15:45 – 17:30
==Track A — LLM Serving & Agents==
#### 4.30

5
study/conf/Untitled.md Normal file
View File

@@ -0,0 +1,5 @@
I am a tax resident of China and my Tax Identification Number is my Chinese Resident Identity Card number.
My current UK address is used as my temporary residential/correspondence address while I am staying in the United Kingdom temporarily. I do not regard the UK as my tax residence and I do not meet the UK tax residence conditions under the Statutory Residence Test.
My permanent home, main personal and financial ties, and tax residence remain in China. I will update HSBC if my tax residence changes in the future.