[Bug]: No error report when passing wrong lora path using num_scheduler_steps=8

[sleepwalker2017]

🐛 Describe the bug

When I run lora request using num_scheduler_steps > 1, even if I give a wrong lora path, it doesn't report an error, but gives wrong result.

The bug can be reproduced by this code.

import os
import torch
from typing import List, Optional, Tuple
import pandas as pd
from huggingface_hub import snapshot_download

from vllm import EngineArgs, LLMEngine, RequestOutput, SamplingParams
from vllm.lora.request import LoRARequest

def create_test_prompts(
        lora_path: str
) -> List[Tuple[str, SamplingParams, Optional[LoRARequest]]]:
    return [
        (
            "What happens when you put oil into water?",  # noqa: E501
            SamplingParams(temperature=1e-7,
                           max_tokens=128,
                           top_k =1,
                           top_p = 1e-5,
                           stop_token_ids=[32003]),
            LoRARequest("sql-lora", 1, lora_path)),
    ]

def process_requests(engine: LLMEngine,
                     test_prompts: List[Tuple[str, SamplingParams,
                                              Optional[LoRARequest]]]):
    """Continuously process a list of prompts and handle the outputs."""
    request_id = 0

    while test_prompts or engine.has_unfinished_requests():
        if test_prompts:
            prompt, sampling_params, lora_request = test_prompts.pop(0)
            engine.add_request(str(request_id),
                               prompt,
                               sampling_params,
                               lora_request=lora_request)
            request_id += 1

        request_outputs: List[RequestOutput] = engine.step()

        for request_output in request_outputs:
            if request_output.finished:
                print(request_output)

def initialize_engine() -> LLMEngine:
    """Initialize the LLMEngine."""
    model_path = "/data/quant_fp8/vicuna-13b-v1.5-fp8/"
    engine_args = EngineArgs(model=model_path,
                             enable_lora=True,
                             max_loras=16,
                             max_lora_rank=64,
                             max_cpu_loras=16,
                             max_num_seqs=256,
                             trust_remote_code=True,
                             enforce_eager=True,
                             num_scheduler_steps=8)
    return LLMEngine.from_engine_args(engine_args)


def main():
    engine = initialize_engine()
    lora_path = '/the/path/does/not/exist/'
    test_prompts = create_test_prompts(lora_path)
    torch.cuda.cudart().cudaProfilerStart()
    process_requests(engine, test_prompts)
    torch.cuda.cudart().cudaProfilerStop()


if __name__ == '__main__':
    main()

Model Input Dumps

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.