Giải mã suy đoán

Normally an LLM generates text one token at a time: each token requires a full forward pass through the giant model, and you can't start the next until the current one finishes. For large models at low batch sizes, decoding can be memory-bandwidth bound: each token step reads many weights, so moving data may dominate arithmetic. Other batch sizes and workloads can have different bottlenecks. Speculative decoding breaks the bottleneck. A small, cheap draft model proposes a chunk of, say, five candidate tokens. The large 'target' model then processes all five in a single parallel forward pass and checks them. In greedy decoding, the target accepts the matching prefix and uses its own token at the first mismatch, discarding the remaining draft. Exact sampling uses a rejection-sampling correction to preserve the target distribution; acceptance is not always simple token equality. The target can verify a drafted block in a parallel pass, but verification still costs computation. Net speedup depends on draft-model cost, how many tokens are accepted, and the hardware and workload.