notes

research and engineering notes on medical AI, scientific computing, safe RL, generative modeling, contrastive learning, and large-model systems

LLM RL: from Bellman Target, PPO Clip to Token/Sequence Ratio and Training-Inference Mismatch

A structured reinforcement learning tutorial from Bellman targets and policy gradients to PPO, GRPO, DAPO, CISPO, GSPO, DPO, and training-inference mismatch.

43 min read · 2026

Orthogonal Polynomials for Uncertainty Quantification: Recurrence Algorithms, PCE, and Biomedical Simulation

A research-level map of orthogonal polynomial recurrence algorithms, polynomial chaos expansion, and noninvasive uncertainty quantification for biomedical simulations.

9 min read · 2025

Contrastive Learning: Objectives, Dictionaries, Momentum Encoders, and Multimodal Alignment

从 anchor/positive/negative、dictionary 与 temperature 出发，梳理 Triplet、NCE/InfoNCE、NT-Xent、MoCo/SimCLR、BYOL/SimSiam/DINO、CLIP 以及 ArcFace/CosFace 的候选集合与表征几何。

36 min read · 2025

a distill-style blog post

an example of a distill-style blog post and main elements

25 min read · 2021

a post with code

an example of a blog post with some code

4 min read · 2015

LLM RL: from Bellman Target, PPO Clip to Token/Sequence Ratio and Training-Inference Mismatch

A structured reinforcement learning tutorial from Bellman targets and policy gradients to PPO, GRPO, DAPO, CISPO, GSPO, DPO, and training-inference mismatch.

43 min read · June 25, 2026

2026 · reinforcement-learning policy-gradient llm-rl ppo rlhf · technical-notes
Inside nano-vLLM: Scheduler, Paged KV Cache, Prefill/Decode, and Sampling

沿请求生命周期拆解 scheduler、paged KV cache、prefill/decode、prefix cache 和采样。

43 min read · April 20, 2026

2026 · vllm inference large-language-models ai-infra · ai-infra
Distributed Training for Large Models: Collectives, ZeRO/FSDP, Tensor, Pipeline, and Expert Parallelism

从 parameter、activation、gradient、optimizer state 的生命周期出发，梳理 All-Reduce/All-Gather/Reduce-Scatter/All-to-All、DDP、ZeRO/FSDP、TP/PP/SP/CP/EP 及显存-通信权衡。

21 min read · March 16, 2026

2026 · distributed-training large-language-models ai-infra deep-learning-systems · ai-infra
Modern Attention for LLMs: MHA/MQA/GQA, RoPE, MLA, FlashAttention, and MoE

从 MHA/MQA/GQA、RoPE、MLA 到 FlashAttention、MoE 的注意力机制笔记。

15 min read · February 22, 2026

2026 · attention transformer large-language-models inference · ai-infra
Flow-Based Generative Models: From Normalizing Flows to Flow Matching, Reflow, and MeanFlow

沿着 exact likelihood、probability path、velocity regression 和 average velocity 这条主线，梳理 NF/CNF、FM/CFM、Gaussian/OT/CondOT path、Stochastic Interpolants、Rectified Flow/Reflow、MeanFlow 与 CFG。

40 min read · January 18, 2026

2026 · flow-matching normalizing-flows generative-models diffusion · ai-notes

notes

research and engineering notes on medical AI, scientific computing, safe RL, generative modeling, contrastive learning, and large-model systems

LLM RL: from Bellman Target, PPO Clip to Token/Sequence Ratio and Training-Inference Mismatch

Orthogonal Polynomials for Uncertainty Quantification: Recurrence Algorithms, PCE, and Biomedical Simulation

Contrastive Learning: Objectives, Dictionaries, Momentum Encoders, and Multimodal Alignment

a distill-style blog post

a post with code

LLM RL: from Bellman Target, PPO Clip to Token/Sequence Ratio and Training-Inference Mismatch

Inside nano-vLLM: Scheduler, Paged KV Cache, Prefill/Decode, and Sampling

Distributed Training for Large Models: Collectives, ZeRO/FSDP, Tensor, Pipeline, and Expert Parallelism

Modern Attention for LLMs: MHA/MQA/GQA, RoPE, MLA, FlashAttention, and MoE

Flow-Based Generative Models: From Normalizing Flows to Flow Matching, Reflow, and MeanFlow