Projects

myTorch

Repository: github.com/syt-nju/myTorch

myTorch is a NumPy-based neural network framework designed to reproduce a streamlined PyTorch-like workflow while keeping the implementation compact and readable.

What I built

• Operator overloading for tensor-style user interaction
• Automatic entry into the computation graph during forward execution
• Graph lifecycle management during backpropagation
• Reusable abstractions built from base classes and decorators

Why it matters

This project strengthened my understanding of code modularity, improved my engineering sense for building reusable abstractions, and gave me a much more concrete grasp of the details behind backpropagation.

Open Source Contributions

verl

• Fixed a bug in the DAPO trainer where confusing step semantics prevented checkpoints from being saved automatically.
• Contributed the fix upstream and merged it into the main branch.

TRL

• Identified and fixed a type-annotation issue in the GRPO trainer.
• Used the contribution process to improve correctness in a commonly used post-training codebase.

What I Like Building

• LLM agents with clear tool-use behavior
• Evaluation systems that make model iteration reliable
• Training and post-training pipelines with strong feedback loops
• Small but well-structured infrastructure for experimentation