| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from typing import NamedTuple |
| 5 | |
| 6 | import openai # use the official client for correctness check |
| 7 | import pytest |
| 8 | import pytest_asyncio |
| 9 | |
| 10 | from tests.utils import RemoteOpenAIServer |
| 11 | from vllm.config import ModelConfig |
| 12 | |
| 13 | # # any model with a chat template should work here |
| 14 | MODEL_NAME = "Qwen/Qwen2-1.5B-Instruct" |
| 15 | |
| 16 | |
| 17 | def get_vocab_size(model_name): |
| 18 | config = ModelConfig( |
| 19 | model=model_name, |
| 20 | seed=0, |
| 21 | dtype="float16", |
| 22 | ) |
| 23 | return config.get_vocab_size() |
| 24 | |
| 25 | |
| 26 | @pytest.fixture(scope="module") |
| 27 | def server(): |
| 28 | args = [ |
| 29 | # use half precision for speed and memory savings in CI environment |
| 30 | "--dtype", |
| 31 | "float16", |
| 32 | "--enforce-eager", |
| 33 | "--max-model-len", |
| 34 | "4080", |
| 35 | "--max-logprobs", # test prompt_logprobs equal to -1 |
| 36 | "151936", |
| 37 | ] |
| 38 | |
| 39 | with RemoteOpenAIServer(MODEL_NAME, args) as remote_server: |
| 40 | yield remote_server |
| 41 | |
| 42 | |
| 43 | @pytest_asyncio.fixture |
| 44 | async def client(server): |
| 45 | async with server.get_async_client() as async_client: |
| 46 | yield async_client |
| 47 | |
| 48 | |
| 49 | class TestCase(NamedTuple): |
| 50 | model_name: str |
| 51 | echo: bool |
| 52 | |
| 53 | |
| 54 | @pytest.mark.asyncio |
| 55 | @pytest.mark.parametrize( |
| 56 | "test_case", |
| 57 | [ |
| 58 | TestCase(model_name=MODEL_NAME, echo=True), |
| 59 | TestCase(model_name=MODEL_NAME, echo=False), |
| 60 | ], |
| 61 | ) |
| 62 | async def test_chat_session_with_echo_and_continue_final_message( |
| 63 | client: openai.AsyncOpenAI, test_case: TestCase |
| 64 | ): |
| 65 | saying: str = "Here is a common saying about apple. An apple a day, keeps" |
| 66 | # test echo with continue_final_message parameter |
| 67 | chat_completion = await client.chat.completions.create( |
| 68 | model=test_case.model_name, |
| 69 | messages=[ |
| 70 | {"role": "user", "content": "tell me a common saying"}, |
| 71 | {"role": "assistant", "content": saying}, |
| 72 | ], |
| 73 | extra_body={ |
| 74 | "echo": test_case.echo, |
| 75 | "continue_final_message": True, |
| 76 | "add_generation_prompt": False, |
| 77 | }, |
| 78 | ) |
| 79 | assert chat_completion.id is not None |
| 80 | assert len(chat_completion.choices) == 1 |
| 81 | |
| 82 | choice = chat_completion.choices[0] |
| 83 | assert choice.finish_reason == "stop" |
| 84 | |
| 85 | message = choice.message |
| 86 | if test_case.echo: |
| 87 | assert message.content is not None and saying in message.content |
| 88 | else: |
| 89 | assert message.content is not None and saying not in message.content |
| 90 | assert message.role == "assistant" |
| 91 | |
| 92 | |
| 93 | @pytest.mark.asyncio |
| 94 | async def test_prompt_logprobs(client: openai.AsyncOpenAI): |
| 95 | messages = [ |
| 96 | {"role": "system", "content": "You are a helpful assistant."}, |
| 97 | {"role": "user", "content": "Beijing is the capital of which country?"}, |
| 98 | ] |
| 99 | |
| 100 | completion = await client.chat.completions.create( |
| 101 | model=MODEL_NAME, |
| 102 | messages=messages, |
| 103 | extra_body={"prompt_logprobs": -1}, |
| 104 | ) |
| 105 | |
| 106 | assert completion.prompt_logprobs is not None |
| 107 | assert len(completion.prompt_logprobs) > 0 |
| 108 | |
| 109 | |
| 110 | @pytest.mark.asyncio |
| 111 | async def test_top_logprobs(client: openai.AsyncOpenAI): |
| 112 | messages = [ |
| 113 | {"role": "system", "content": "You are a helpful assistant."}, |
| 114 | {"role": "user", "content": "Beijing is the capital of which country?"}, |
| 115 | ] |
| 116 | |
| 117 | completion = await client.chat.completions.create( |
| 118 | model=MODEL_NAME, |
| 119 | messages=messages, |
| 120 | max_tokens=1, |
| 121 | extra_body={ |
| 122 | "top_logprobs": -1, |
| 123 | "logprobs": "true", |
| 124 | }, |
| 125 | ) |
| 126 | assert completion.choices[0].logprobs is not None |
| 127 | assert completion.choices[0].logprobs.content is not None |
| 128 | assert len(completion.choices[0].logprobs.content) > 0 |
| 129 | assert len( |
| 130 | completion.choices[0].logprobs.content[0].top_logprobs |
| 131 | ) == get_vocab_size(MODEL_NAME) |