Blog¶
Technical notes on machine learning, GPU programming, and security.
Notes on Serving LLMs with TensorRT-LLM and Triton
From a Hugging Face checkpoint to a production TensorRT-LLM + Triton endpoint, benchmarked against vLLM.
2026-05-31
Where tensor-parallel inference hits the NVLink wall
Measuring NCCL all-reduce on NVLink and why tensor-parallel decode hits a small-message latency wall, not a bandwidth one.
2026-05-31
0 % vs 40 %: making a RAG agent refuse to hallucinate
One guardrail that drops out-of-corpus hallucination from ~40% to 0% — and why an eval harness is the only way to see it.
2026-05-31
Notes on Federated Learning and Differential Privacy
FedAvg / FedProx / SCAFFOLD from scratch, what breaks under Non-IID data, and how DP-SGD adds real guarantees.
2026-05-31
Notes on CUDA Tensor Core GEMM (WMMA)
Matrix multiply from naive to tiled to WMMA — shared-memory tiling, fragments, and the cuBLAS ceiling.
2026-05-31
Meta-Reinforcement Learning of Structured Exploration Strategies
「學習如何學習」— how meta-RL lets an agent adapt quickly to new environments and tasks.
2025-01-11
CUDA Programming 入門
NVIDIA 平行運算平台與程式設計模型入門:threads、blocks、記憶體階層。
2025-01-11
滲透測試入門技術
滲透測試的定義、流程、類型與工具 — 模擬駭客攻擊找出可修補的漏洞。
2025-01-10