Managing modules at a production level in Haskell with Cabal or Stack.

Advanced topic: GADT in Haskell, it was also called Phantom Types

Roofline model is a theoretical model to analyze whether a kernel is bounded by computation or memory.

OpenAI released tool recently, built on Triton to fit Blackwell's warp-based execution model.

Accepted by ICLR 2026. To improve LLMs' ability on multi-step retrieval, instead of fine-tuning LLMs, a method that use RL to train embedding models is proposed.

Avalanche is a protocol for Blockchain, and we may learn about L1 Design from Avalanche.

联邦学习领域中一种保护隐私的理论

Just a simple diary on running swe-bench, a classic agentic task benchmark.

D 我们直接枚举 3n3^n3n 个 mask (i.e., 0,1,?0,1,\texttt{?}0,1,?) 考虑 mask 的第 nnn 位，如果是 0,10,10,1，那么剩下的其实就是大小为 n−1n-1n−1 的一个子问题；比较麻烦的是...

Learn (1) integrate CUDA kernels into PyTorch (2) profiling CUDA kernels in PyTorch