After a bunch of tests:

• pytest tests/test_vllm_client_server.py for client-server
• for colocate:

  from datasets import load_dataset
  from trl import GRPOTrainer, GRPOConfig
  from trl.rewards import accuracy_reward
  
  dataset = load_dataset("trl-lib/DeepMath-103K", split="train[:100]")
  
  trainer = GRPOTrainer(
      model="Qwen/Qwen3-0.6B",
      reward_funcs=accuracy_reward,
      args=GRPOConfig(use_vllm=True, vllm_mode="colocate"),
      train_dataset=dataset,
  )
  trainer.train()

it seems like TRL works with vLLM 0.11.x