用 Little 定律解释推测解码在真实服务中的提速曲线 —— 阅读笔记
Kong 等人提出的「面向真实服务的推测解码延迟模型」阅读笔记。用 roofline 风格的延迟分解加 Little 定律，把不同 RPS、模型、硬件下的延迟曲线压缩到同一条 1/(1-x) 通用形上，并从机制层面解释了「batch=1 SD 提速在高负载下消失」的现象。
An Interpretable Latency Model for Speculative Decoding in LLM Serving — Technical Review
Posted on
In
LLM Inference
,
Serving Systems
A detailed technical review of Kong et al.'s interpretable latency model for speculative decoding under real serving workloads. Using a roofline-style decomposition plus Little's Law, the paper collapses RPS-versus-latency curves onto a single universal form and gives a mechanistic explanation for why batch=1 SD speedups erode under load.
A detailed technical review of Kong et al.'s interpretable latency model for speculative decoding under real serving workloads. Using a roofline-style decomposition plus Little's Law, the paper collapses RPS-versus-latency curves onto a single universal form and gives a mechanistic explanation for why batch=1 SD speedups erode under load.
Zero Sum SVD：用「损失零和」做全局奇异值预算分配的 LLM 压缩方法
Posted on
In
LLM Compression
,
Model Optimization
一篇关于 Zero Sum SVD 的中文阅读笔记：把所有层的奇异值堆到一个全局优先队列里，用带符号的损失敏感度和「零和守恒」的贪心规则一次性决定全模型的秩预算，让异质化的逐层秩自然从一条标量约束里掉出来。
一篇关于 Zero Sum SVD 的中文阅读笔记：把所有层的奇异值堆到一个全局优先队列里，用带符号的损失敏感度和「零和守恒」的贪心规则一次性决定全模型的秩预算，让异质化的逐层秩自然从一条标量约束里掉出来。
Zero Sum SVD: A Global, Loss-Aware Rank Budget for LLM Compression
Posted on
In
LLM Compression
,
Model Optimization
A detailed technical review of Zero Sum SVD, which replaces per-layer rank optimization with a global, signed loss-sensitivity heap and a greedy zero-sum rule, letting heterogeneous per-layer ranks fall out of one scalar conservation law.
A detailed technical review of Zero Sum SVD, which replaces per-layer rank optimization with a global, signed loss-sensitivity heap and a greedy zero-sum rule, letting heterogeneous per-layer ranks fall out of one scalar conservation law.
DisagMoE：用解耦 Attention 和 FFN 打通 MoE 训练的 all-to-all 瓶颈
Posted on
Edited on
In
LLM Training
,
Distributed Systems
一篇关于 DisagMoE 的中文阅读笔记：把 attention 和 FFN 分别放到独立 GPU 池，用 AF-Pipe 调度和 M2N 通讯原语把两侧拼起来，从而把 MoE 训练里的 all-to-all 瓶颈藏进计算之下。
一篇关于 DisagMoE 的中文阅读笔记：把 attention 和 FFN 分别放到独立 GPU 池，用 AF-Pipe 调度和 M2N 通讯原语把两侧拼起来，从而把 MoE 训练里的 all-to-all 瓶颈藏进计算之下。
DisagMoE: Disaggregating Attention and FFN to Beat the MoE All-to-All Bottleneck
Posted on
Edited on
In
LLM Training
,
Distributed Systems
A detailed technical review of DisagMoE, which disaggregates attention and FFN layers onto separate GPU pools and stitches them together via the AF-Pipe schedule to hide the MoE all-to-all bottleneck during training.
A detailed technical review of DisagMoE, which disaggregates attention and FFN layers onto separate GPU pools and stitches them together via the AF-Pipe schedule to hide the MoE all-to-all bottleneck during training.
DAPO：大规模开源 LLM 强化学习系统
Posted on
In
RL Training
,
LLM Training
一篇关于 DAPO 的中文阅读笔记：它把 Clip-Higher、动态采样、token-level loss 与 overlong reward shaping 组合成可复现的大规模 LLM 强化学习配方。
一篇关于 DAPO 的中文阅读笔记：它把 Clip-Higher、动态采样、token-level loss 与 overlong reward shaping 组合成可复现的大规模 LLM 强化学习配方。
DAPO: An Open-Source LLM Reinforcement Learning System at Scale
Posted on
In
RL Training
,
LLM Training
A detailed technical review of DAPO, an open-source large-scale reinforcement learning recipe for reasoning LLMs using Clip-Higher, dynamic sampling, token-level loss, and overlong reward shaping.
A detailed technical review of DAPO, an open-source large-scale reinforcement learning recipe for reasoning LLMs using Clip-Higher, dynamic sampling, token-level loss, and overlong reward shaping.
MASPO：面向 LLM 多智能体系统的联合提示词优化
Posted on
In
AI Agents
,
Prompt Optimization
一篇关于 MASPO 的中文阅读笔记：它用 local、lookahead 与 global 三类信号联合优化 LLM 多智能体系统中的角色提示词。
一篇关于 MASPO 的中文阅读笔记：它用 local、lookahead 与 global 三类信号联合优化 LLM 多智能体系统中的角色提示词。
MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems
Posted on
In
AI Agents
,
Prompt Optimization
A detailed technical review of MASPO, a joint prompt optimization method for multi-agent LLM systems that balances local, downstream, and global rewards.
A detailed technical review of MASPO, a joint prompt optimization method for multi-agent LLM systems that balances local, downstream, and global rewards.
Tutti：让基于 SSD 的 KV Cache 真正适用于长上下文 LLM Serving
Posted on
Edited on
In
ML Systems
,
LLM Inference
一篇关于 Tutti 的中文阅读笔记：它从 GPU-native KV cache object store、GPU io_uring 与 slack-aware scheduling 出发，让 SSD-backed KV cache 更适合长上下文 LLM serving。
一篇关于 Tutti 的中文阅读笔记：它从 GPU-native KV cache object store、GPU io_uring 与 slack-aware scheduling 出发，让 SSD-backed KV cache 更适合长上下文 LLM serving。
Tutti: Making SSD-Backed KV Cache Practical for Long-Context LLM Serving
Posted on
Edited on
In
ML Systems
,
LLM Inference
A detailed technical review of Tutti, a GPU-centric SSD-backed KV cache system that makes long-context LLM serving cache reuse practical.
A detailed technical review of Tutti, a GPU-centric SSD-backed KV cache system that makes long-context LLM serving cache reuse practical.