LeetCode for people who'd rather train models than grind arrays

What it does TorchLeet is a curated set of Jupyter notebook exercises that treat PyTorch like a language you learn by speaking, not reading. Problems run from basic (custom loss functions, TensorBoard logging) through hard (custom autograd ops, GANs, Vision Transformers) and into a newer LLM track that covers attention mechanisms, ROPE embeddings, and even speculative decoding. Each problem ships as an incomplete notebook with ... gaps and #TODO comments; a matching solution notebook sits nearby for post-attempt comparison.

The interesting bit The author admits to using GPT to generate the problem set, then solving them manually as practice—a refreshingly honest workflow that doubles as quality control. The LLM track is the real differentiator: it treats modern transformer internals as learnable primitives rather than black-box imports, asking you to implement KV cache, grouped query attention, and a full SmolLM from scratch.

Key highlights

• ~40+ problems across four difficulty tiers plus a dedicated LLM set
• Heavy emphasis on “from scratch” implementations (CNNs, LSTMs, byte-pair encoding, 5D parallelism)
• Incomplete notebook format forces active coding rather than passive reading
• Covers interview-relevant territory: quantization, LoRA/QLoRA, RLHF/DPO, continuous batching
• Open for community contributions with a defined problem/solution structure

Caveats

• Several listed problems (Neural Style Transfer, GNN/GCN, AlexNet, sparse tensors, distributed training) appear as text-only entries without linked notebooks
• The README’s structure description mentions questions/ and solutions/ directories that don’t cleanly match the actual torch/ and llm/ paths shown
• Some LLM problems are similarly text-only (KL divergence, RMS Norm, GPTQ, various sampling methods)

Verdict Solid fit for developers who know ML concepts but need to translate them into raw PyTorch under interview pressure. Less useful if you want a guided course with prose explanations—the value is in the blank-space forcing function.