How to assign a separate GPU to each worker?

[Dawnfz-Lenfeng]

I am trying to use Bentoml to run the vllm engine and set up 4 workers, but it seems they are all running on a single GPU. I want to assign each worker to a separate GPU. How can I configure this?

@openai_endpoints(
    model_id="llama-3-instruct",
    default_chat_completion_parameters=dict(stop=["<|eot_id|>"]),
)
@bentoml.service(
    name="bentovllm-llama3-8b-insruct-service",
    traffic={
        "timeout": 300,
        "concurrency": 100,  # Matches the default max_num_seqs in the VLLM engine
    },
    resources={
        "gpu": 4,
    },
    workers=4,
    workers_per_resource=1
)
class VLLM:
    def __init__(self) -> None:
        ENGINE_ARGS = AsyncEngineArgs(
            model=MODEL_ID,
            served_model_name="llama-3-instruct",
            # tensor_parallel_size=1,
            gpu_memory_utilization=0.85,
            max_num_seqs=100,
        )

        self.engine = AsyncLLMEngine.from_engine_args(ENGINE_ARGS)

        tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
        self.stop_token_ids = [
            tokenizer.eos_token_id,
            tokenizer.convert_tokens_to_ids("<|eot_id|>"),
        ]