torchcode-pytorch-interview-practice — Claude Code skill

What torchcode-pytorch-interview-practice does

The skill walks you through reimplementing PyTorch layers from scratch: softmax, LayerNorm, MultiHeadAttention, GPT-2 and more. Problems are graded easy, medium and hard. Each one lives in its own notebook, with a blank template to work in and a reference solution to read once you have tried it yourself.

Grading comes from the torch_judge package. Calling check("softmax") runs the test cases, compares your output against the PyTorch reference, verifies autograd compatibility and reports timing. hint() gives a nudge without spoiling the answer, reset_progress() clears a problem, and status() lists the problems with your progress.

You can skip installation entirely: the notebooks open in Hugging Face Spaces and Google Colab. To run them in your own environment there is pip install torch-judge, a prebuilt Docker image, or a local build from the repository via make run. That last one detects Docker or Podman and falls back to a local build when the registry image is unavailable, which happens often on Apple Silicon.