vLLM vllm/v1/sample/rejection_sampler.py (main)

The verify step's per-token cost as shipped. The target model's logits arrive as ONE [num_tokens + batch_size, vocab_size] tensor, are float()-cast, then INDEXED TWICE by tensor (bonus_logits_indices and target_logits_indices); PyTorch gives each index its own storage, so the [.., V] tensor is copied twice. The target slice is .to(torch.float32) and then, whenever logprobs are not requested in processed mode, .clone()d so the raw copy survives the in-place logit processors - a third full-width fp32 buffer. draft_probs is an optional [num_tokens, vocab_size] tensor (absent for ngram spec decode), and the registered kernels are separate ops: _rejection_greedy_sample, _rejection_random_sample, _expand, _sample_recovered_tokens, plus synthetic acceptance-rate and fp64-Gumbel variants. So rejection sampling's per-token cost is k x V fp32 written and re-read several times, and it is measured in extra kernels, not extra math.