Implement Speculative Decoding in PyTorch

Hardllm-inferenceauto-graded

Draft tokens with a fast model, verify in parallel with the target model, and accept/reject to guarantee identical output distribution.

Solve it

Check your answer

The grader verifies properties of your implementation, so a correct solution written differently from ours still passes.

pip install torchleet

from torchleet import check
check("speculative-decoding", DraftModel, TargetModel, standard_decode, speculative_decode)

Company tags

How these tags are sourced