Codeing, AI, anime, games, and the occasional journey into a galgame world.
- 💻 I love coding, especially building things with C++ and understanding how systems work under the hood.
- 🤖 I'm exploring AI Infrastructure, CUDA, LLM inference, and the internals hidden behind frameworks such as PyTorch.
- 🧩 I enjoy learning by rebuilding: tensors, CUDA operators, tokenizers, model loading, KV cache, and autoregressive generation.
- 📺 Away from the terminal, you'll usually find me watching anime.
- 🎮 I also enjoy games, especially story-rich games and galgames / visual novels.
- 🌙 My ideal rhythm: write code, learn something new about AI, then disappear into a fictional world for a while.
Build systems by day, explore fictional worlds by night.
An educational Qwen3 Dense inference runtime built from scratch with C++17 and CUDA. It opens the framework black box and reconstructs the inference path piece by piece:
- custom tensor, buffer, dtype, shape, and device abstractions;
- memory-mapped and sharded SafeTensors loading;
- a custom byte-level BPE tokenizer;
- CUDA operators, cuBLAS linear layers, RoPE, GQA, and RMSNorm;
- KV cache, sampling, autoregressive generation, and terminal chat;
- support for architecture-compatible models in the Qwen3 Dense family.
The project favors clarity and learning first, then gradually moves toward profiling and performance optimization.
| 💻 Coding | 🤖 AI | 📺 Anime | 🎮 Games | 💗 Galgame |
|---|---|---|---|---|
| Build | Explore | Watch | Play | Experience stories |
Good code is satisfying. A good story is unforgettable. I like worlds built from both.
The language card reflects the composition of public repositories, not a ranking of proficiency.