| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import pytest |
| 5 | import torch.cuda |
| 6 | |
| 7 | from vllm import LLM, SamplingParams |
| 8 | from vllm.platforms import current_platform |
| 9 | from vllm.v1.engine import EngineCoreRequest |
| 10 | from vllm.v1.engine.core import EngineCore |
| 11 | |
| 12 | MODEL_NAME = "hmellor/tiny-random-LlamaForCausalLM" |
| 13 | |
| 14 | |
| 15 | def test_preprocess_error_handling(monkeypatch: pytest.MonkeyPatch): |
| 16 | """Test that preprocessing errors are handled gracefully.""" |
| 17 | if current_platform.is_rocm() or current_platform.is_xpu(): |
| 18 | pytest.skip( |
| 19 | "Skipped on ROCm/XPU: this test only works with 'fork', " |
| 20 | "but ROCm/XPU uses 'spawn'." |
| 21 | ) |
| 22 | |
| 23 | assert not torch.cuda.is_initialized(), ( |
| 24 | "fork needs to be used for the engine " |
| 25 | "core process and this isn't possible if cuda is already initialized" |
| 26 | ) |
| 27 | |
| 28 | # Store original method to call for non-failing requests |
| 29 | original_preprocess = EngineCore.preprocess_add_request |
| 30 | |
| 31 | # Monkeypatch to make preprocess_add_request raise an exception |
| 32 | # only for requests with "FAIL" in the first token |
| 33 | def conditional_failing_preprocess(self, request: EngineCoreRequest): |
| 34 | # Fail if the first token id is 333 |
| 35 | if request.prompt_token_ids and request.prompt_token_ids[0] == 333: |
| 36 | raise ValueError("Simulated preprocessing error!") |
| 37 | return original_preprocess(self, request) |
| 38 | |
| 39 | monkeypatch.setattr( |
| 40 | EngineCore, "preprocess_add_request", conditional_failing_preprocess |
| 41 | ) |
| 42 | |
| 43 | llm = LLM(model=MODEL_NAME) |
| 44 | |
| 45 | # Create a failing request by crafting a request with an invalid token |
| 46 | # We need to use a direct approach since LLM.generate tokenizes for us |
| 47 | from vllm.inputs import TokensPrompt |
| 48 | |
| 49 | # This should raise an exception due to the preprocessing failure |
| 50 | # Special token id to trigger the failure |
| 51 | failing_prompt = TokensPrompt(prompt_token_ids=[333]) |
| 52 | outputs = llm.generate(failing_prompt, SamplingParams(max_tokens=10)) # type: ignore |
| 53 | assert len(outputs) == 1 |
| 54 | assert len(outputs[0].outputs[0].token_ids) == 0 |
| 55 | assert outputs[0].finished |
| 56 | assert outputs[0].outputs[0].finish_reason == "error" |
| 57 | |
| 58 | # Verify the engine is still functional with a normal request |
| 59 | outputs = llm.generate("Hello, my name is", SamplingParams(max_tokens=10)) |
| 60 | assert len(outputs) == 1 |
| 61 | assert len(outputs[0].outputs[0].token_ids) > 0 |
| 62 | assert outputs[0].outputs[0].finish_reason in ("stop", "length") |