Source ReadingNo. 001
SkyPilot
Cloud orchestration as an optimization problem: resources, regions, spot instances, Kubernetes, jobs, and the control loops that make them practical.
The Library · 文库
The Library
一个分门别类的索引， 收齐了我写过的东西。 代码精读和论文精读是 /sources 下自成一体的 HTML 深读； 教程和博客是中英双语的 blog 长文； 实验是可复现的 kernel 与系统性能记录， 连原始数据一起存档。 五个栏目， 一张地图。
11code readings
5paper readings
7tutorials
6blog posts
3experiments
01
Code
代码精读 · source-level readings of real codebases — each a self-contained HTML deep dive
11 deep dives
Source ReadingNo. 002
SGLang
A serving-system read focused on runtime structure, request scheduling, cache management, routing, and the boundary between Python control and fast inference paths.
Source ReadingNo. 003
vLLM
The wrap-up of the initial serving trilogy: scheduler pressure, PagedAttention, KV memory, batching, and how design choices compare against SGLang.
Source ReadingNo. 004
mini-SGLang
A smaller codebase read as a teaching artifact: what a minimal implementation makes explicit, what it hides, and how to learn from that compression.
Source ReadingNo. 005
GCNasm
A descent into hand-written AMD GPU assembly: CDNA3 idioms, occupancy, memory movement, instruction selection, and the optimization patterns behind fast kernels.
Source ReadingNo. 006
FlyDSL
A Python DSL with typed MLIR underneath: layout algebra, copy and MMA atoms, compiler boundaries, and what it takes to express production GEMM from Python.
Source ReadingNo. 007
AITER MoE Tuner
A tuner-first reading of MoE GEMM search: config space, benchmarking discipline, hardware assumptions, and why tuning code is often kernel knowledge in disguise.
Source ReadingNo. 008
rocprof Viewer
A field guide to AMD instruction-level profiling: rocprofv3 capture, Advanced Thread Trace, source mapping, and how to read the viewer panels without fooling yourself.
Source ReadingNo. 009
Codex Goal Mode
A source-level reading of goal mode as a thread-scoped state machine: persisted goals, model tools, runtime continuation, token budget accounting, and authority boundaries.
Source ReadingNo. 010
AITER CI Kernel Surface
A source-level map of AITER's active CI: workflow routing, kernel families, shape pressure points, and the downstream gates that turn kernel tests into system tests.
Source ReadingNo. 011
Humanize
A Claude Code plugin that wraps an agent in a shell-hook control loop with an independent Codex reviewer: the phase is a filename, a dozen cheap gates guard one expensive review, and a feedback controller trips circuit breakers when progress stalls.
02
Paper
论文精读 · close readings of research papers, rebuilt as HTML deep dives
5 deep dives
Paper ReadingNo. 005
Draft-OPD
A speculative-decoding draft model trained by plain SFT plateaus — it never sees its own mistakes. Draft-OPD teaches it from its rejected guesses: target-assisted rollout, error-position replay, and an acceptance-aware forward/reverse-KL objective. +23% over EAGLE-3, +13% over DFlash, with every number checked against the source.
Paper ReadingNo. 004
The Chunked Pipeline
How SGLang serves million-token prompts with pipeline parallelism: the communication-volume math, the bubble ratio, async P2P, and the dynamic chunking that equalizes per-stage time. A first-principles rebuild of the LMSYS post.
Paper ReadingNo. 001
Polar
Agentic RL without rewriting the harness: proxying LLM API calls, asynchronous staging, prefix merging, and what SWE-Bench tells us about scalable agent training.
Paper ReadingNo. 002
Kernel Design Agents
A close read of agentic GPU kernel development: plan-execute-verify loops, KernelWiki, ncu-guided debugging, autotuning, and reward-hacking failure modes.
Paper ReadingNo. 003
Linear Layouts
One binary matrix over GF(2) as the organizing principle for tensor layouts: conversion, broadcast, swizzling, slicing, and robust code generation.
03
Tutorial
教程 · first-principles primers and guides — one rich HTML primer, the rest bilingual blog posts
7 pieces
Guide · HTMLkernel autotune
Autotune, End to End
How kernel autotuning is designed in Triton, used in aiter / quack / CuteDSL, and consumed by engines like SGLang — and from all of it, how to design FlyDSL's autotune, with a concrete implementation guide.
Primer · HTMLcompiler stack
From Python to Silicon
A systems primer for the path from Python to GPU execution: compiler layers, kernel boundaries, IR, runtime dispatch, and what each layer is responsible for.
Primerattention
Attention Mechanisms
Full, Sparse, and Linear attention from first principles — up through DeepSeek NSA and Gated Linear Attention, with the tradeoffs that decide each one.
Primerinference
KV Cache & Model Weights
The first thing to understand before optimizing inference: what KV cache is, how it differs from model weights, and how each scales with sequence and batch.
PrimerGPU memory
LLM GPU Memory Calculation
How to actually compute LLM memory on a GPU — the components, worked 7B/70B examples, and how DP / TP / PP / EP and ZeRO change the arithmetic.
Guidepost-training
SFT & RL Training Guide
A first-principles guide to SFT and RL post-training: loss and label masking, dataset construction, hyperparameters, RLHF, and the common pitfalls.
Primertransformer
Transformer Deep Dive
The Transformer rebuilt from three angles at once — the math, runnable PyTorch, and the design rationale behind self-attention, LayerNorm, and the MLP.
04
Blog
博客 · original writing — benchmarks, framework comparisons, and project notes
6 posts
Benchmarklaunch tuning
Kimi-K3 × DSpark on MI355X
A 450-minute launch-parameter search plus an 18-point characterisation of the winning recipe — why every throughput decision reduces to accepted tokens over step latency.
Benchmarkspec decoding
Qwen3-Coder × EAGLE3
A measured benchmark of EAGLE3 speculative decoding on Qwen3-Coder-30B-A3B — where the 1.87× speedup comes from and why code generation benefits most.
ComparisonRL frameworks
NeMo-RL vs slime
A working comparison of two RL post-training frameworks — algorithms, engineering quality, MoE support, and ROCm fit — with a reasoned pick for MI300X / MI355X.
ProjectRL · kernels
TritonForge
Building a server-based, multi-turn RL system that generates Triton kernels across NVIDIA and AMD — architecture, SFT+RL methodology, results, and roadmap.
NoteFlyDSL
BasisAttr · beneath Layout
A follow-up note beneath the FlyDSL layout algebra: what BasisAttr and Fly_Basis are, why layouts need them, and where to start completing the surface.
AnalysisMoE RL · long-context
Long-Sequence MoE RL
A first-principles read of Yan Bai's 128K-context MoE RL recipe — Path B recompute, linear cross-entropy, FSDP2, chunked EP overlap — re-derived, fact-checked, and mapped onto MI300X / MI355X.
05
Experiments
实验记录 · reproducible kernel and system measurements — raw data archived alongside the write-up
3 records
ExperimentNo. 003
Faster kernels, slower engine: TokenSpeed vs SGLang on Kimi-K3
TokenSpeed uses 17.6% less non-collective device time in graph-off traces, yet is 1.14x slower in graph-on batch-1 serving and about 3.7x slower at concurrency 8/32. Pinned code links and three follow-up experiments separate what the data proves from the leading host, scheduler and FlatKV hypotheses.
ExperimentNo. 002
Kimi-K3: three of every four layers are linear
Attributing decode and prefill GPU time to KDA, full attention and MoE across 1K–1M context on 8x MI355X — 74% of the layers turn out to cost 9–24% of the time, and a KDA layer is only the cheaper of the two above ~4K decode / ~8K prefill.
ExperimentNo. 001
Quack RMSNorm on H100 & H200
A tuned and un-tuned sweep of Quack's CuTe DSL RMSNorm forward and backward kernels across 90 cells on two idle Hopper GPUs — and why the analytical config ladder gives up 17% on the widest backward row.
Code, Paper & Experiments open as self-contained HTML deep dives (each carries its own EN / ZH toggle). Tutorial & Blog open as bilingual blog pages. Search across everything, or filter by shelf. Try: MLIR, ATT, MoE, PagedAttention, agent RL, GF(2), RLHF, Codex, RMSNorm.