Hugging Face

Speculative Decoding in PyTorch: Draft-and-Verify Sampling

Members only

An onsite ML coding round where you build a core LLM inference speedup end to end in PyTorch on real HuggingFace models. It rewards candidates who can write working generation code and also defend the probability math behind it. The hard part is getting both the tensor bookkeeping and the correctness argument right under interview pressure.