[Core] Performance: Use list[np.ndarray] instead of list[list[int]] for output tokens for GC optimization (#26368)

Signed-off-by: Jialin Ouyang <Jialin.Ouyang@gmail.com>

tests/v1/kv_connector/unit/utils.py

@@ -7,6 +7,7 @@ from dataclasses import dataclass
 from itertools import chain, count
 from typing import Any
 
+import numpy as np
 import torch
 
 from vllm import SamplingParams
@@ -228,7 +229,7 @@ def create_model_runner_output(
 
     # Make sampled tokens.
     sampled_token = EOS_TOKEN_ID if use_eos else token_id
-    sampled_token_ids = [[sampled_token] for _ in req_ids]
+    sampled_token_ids = [np.array([sampled_token]) for _ in req_ids]
 
     kv_connector_output = (
         None