diff --git a/vllm/v1/spec_decode/llm_base_proposer.py b/vllm/v1/spec_decode/llm_base_proposer.py index 1c59704de6..8d38038734 100644 --- a/vllm/v1/spec_decode/llm_base_proposer.py +++ b/vllm/v1/spec_decode/llm_base_proposer.py @@ -883,16 +883,12 @@ class SpecDecodeBaseProposer: is not sampled and comes from `request.get_token_id()` instead. This is denoted the "backup" token id. It also counts rejected tokens via `sampled_token_ids`. """ - # Precompute get_token_id for when there is no valid next token + # Precompute backup token IDs for discarded requests. num_reqs = gpu_input_batch.num_reqs - seq_lens_list = (gpu_input_batch.num_tokens_no_spec[:num_reqs] - 1).tolist() - self.backup_next_token_ids.np[:num_reqs] = np.array( - [ - requests[gpu_input_batch.req_ids[i]].get_token_id(seq_lens_list[i]) - for i in range(num_reqs) - ], - dtype=np.int32, - ) + for i in range(num_reqs): + self.backup_next_token_ids.np[i] = requests[ + gpu_input_batch.req_ids[i] + ].get_token_id(gpu_input_batch.num_tokens_no_spec[i] - 1) self.backup_next_token_ids.copy_to_gpu(num_reqs) backup_tokens_gpu = self.backup_next_token_ids.gpu