| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Test error handling in Processor. Should not impact other reqs.""" |
| 4 | |
| 5 | import asyncio |
| 6 | |
| 7 | import pytest |
| 8 | |
| 9 | from tests.v1.shutdown.utils import SHUTDOWN_TEST_TIMEOUT_SEC |
| 10 | from vllm import SamplingParams |
| 11 | from vllm.engine.arg_utils import AsyncEngineArgs |
| 12 | from vllm.inputs import ExplicitEncoderDecoderPrompt |
| 13 | from vllm.sampling_params import RequestOutputKind |
| 14 | from vllm.v1.engine.async_llm import AsyncLLM |
| 15 | from vllm.v1.engine.exceptions import EngineGenerateError |
| 16 | |
| 17 | MODELS = ["meta-llama/Llama-3.2-1B"] |
| 18 | |
| 19 | |
| 20 | @pytest.mark.asyncio |
| 21 | @pytest.mark.timeout(SHUTDOWN_TEST_TIMEOUT_SEC) |
| 22 | @pytest.mark.parametrize("model", MODELS) |
| 23 | async def test_async_llm_processor_error(model: str) -> None: |
| 24 | """Test that AsyncLLM propagates a processor error. |
| 25 | Test empty tokens prompt (failure) and non-empty prompt (no failure.) |
| 26 | AsyncLLM always uses an MP client. |
| 27 | """ |
| 28 | engine_args = AsyncEngineArgs(model=model, enforce_eager=True) |
| 29 | async_llm = AsyncLLM.from_engine_args(engine_args) |
| 30 | |
| 31 | async def generate(request_id: str): |
| 32 | # An encoder/decoder prompt is rejected by the Processor for a |
| 33 | # decoder only model. |
| 34 | generator = async_llm.generate( |
| 35 | ExplicitEncoderDecoderPrompt( |
| 36 | encoder_prompt="Hello my name is", decoder_prompt=None |
| 37 | ), |
| 38 | request_id=request_id, |
| 39 | sampling_params=SamplingParams(), |
| 40 | ) |
| 41 | try: |
| 42 | async for _ in generator: |
| 43 | pass |
| 44 | except Exception as e: |
| 45 | return e |
| 46 | |
| 47 | NUM_REQS = 3 |
| 48 | tasks = [generate(f"request-{idx}") for idx in range(NUM_REQS)] |
| 49 | outputs = await asyncio.gather(*tasks) |
| 50 | |
| 51 | # Every request should have get an EngineGenerateError. |
| 52 | for output in outputs: |
| 53 | with pytest.raises(EngineGenerateError): |
| 54 | raise output |
| 55 | |
| 56 | # AsyncLLM should be errored. |
| 57 | assert not async_llm.errored |
| 58 | |
| 59 | # This should be no problem. |
| 60 | EXPECTED_TOKENS = 5 |
| 61 | outputs = [] |
| 62 | async for out in async_llm.generate( |
| 63 | "Hello my name is", |
| 64 | request_id="abc", |
| 65 | sampling_params=SamplingParams( |
| 66 | max_tokens=EXPECTED_TOKENS, output_kind=RequestOutputKind.DELTA |
| 67 | ), |
| 68 | ): |
| 69 | outputs.append(out) |
| 70 | |
| 71 | generated_tokens = [] |
| 72 | for out in outputs: |
| 73 | generated_tokens.extend(out.outputs[0].token_ids) |
| 74 | assert len(generated_tokens) == EXPECTED_TOKENS |
| 75 | |
| 76 | async_llm.shutdown() |