| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import weakref |
| 5 | |
| 6 | import pytest |
| 7 | |
| 8 | from vllm import LLM, SamplingParams |
| 9 | from vllm.exceptions import VLLMValidationError |
| 10 | |
| 11 | MODEL_NAME = "distilbert/distilgpt2" |
| 12 | |
| 13 | PROMPTS = [ |
| 14 | "Hello, my name is", |
| 15 | "The president of the United States is", |
| 16 | "The capital of France is", |
| 17 | "The future of AI is", |
| 18 | ] |
| 19 | |
| 20 | TOKEN_IDS = [ |
| 21 | [0], |
| 22 | [0, 1], |
| 23 | [0, 2, 1], |
| 24 | [0, 3, 1, 2], |
| 25 | ] |
| 26 | |
| 27 | |
| 28 | @pytest.fixture(scope="module") |
| 29 | def llm(vllm_runner): |
| 30 | with vllm_runner( |
| 31 | MODEL_NAME, |
| 32 | max_num_batched_tokens=4096, |
| 33 | tensor_parallel_size=1, |
| 34 | gpu_memory_utilization=0.10, |
| 35 | enforce_eager=True, |
| 36 | ) as runner: |
| 37 | # pytest caches yielded fixtures until after teardown, so use a proxy to |
| 38 | # avoid retaining the LLM while VllmRunner.__exit__ releases ROCm memory. |
| 39 | yield weakref.proxy(runner.llm) |
| 40 | |
| 41 | |
| 42 | @pytest.mark.skip_global_cleanup |
| 43 | def test_multiple_sampling_params(llm: LLM): |
| 44 | sampling_params = [ |
| 45 | SamplingParams(temperature=0.01, top_p=0.95), |
| 46 | SamplingParams(temperature=0.3, top_p=0.95), |
| 47 | SamplingParams(temperature=0.7, top_p=0.95), |
| 48 | SamplingParams(temperature=0.99, top_p=0.95), |
| 49 | ] |
| 50 | |
| 51 | # Multiple SamplingParams should be matched with each prompt |
| 52 | outputs = llm.generate(PROMPTS, sampling_params=sampling_params) |
| 53 | assert len(PROMPTS) == len(outputs) |
| 54 | |
| 55 | # Exception raised, if the size of params does not match the size of prompts |
| 56 | with pytest.raises(VLLMValidationError): |
| 57 | outputs = llm.generate(PROMPTS, sampling_params=sampling_params[:3]) |
| 58 | |
| 59 | # Single SamplingParams should be applied to every prompt |
| 60 | single_sampling_params = SamplingParams(temperature=0.3, top_p=0.95) |
| 61 | outputs = llm.generate(PROMPTS, sampling_params=single_sampling_params) |
| 62 | assert len(PROMPTS) == len(outputs) |
| 63 | |
| 64 | # sampling_params is None, default params should be applied |
| 65 | outputs = llm.generate(PROMPTS, sampling_params=None) |
| 66 | assert len(PROMPTS) == len(outputs) |
| 67 | |
| 68 | |
| 69 | def test_multiple_priority(llm: LLM): |
| 70 | # Generate works when priority is None |
| 71 | outputs = llm.generate(PROMPTS, sampling_params=None, priority=None) |
| 72 | assert len(PROMPTS) == len(outputs) |
| 73 | |
| 74 | # Generate works when length of priority is same as the len(PROMPTS) |
| 75 | outputs = llm.generate(PROMPTS, sampling_params=None, priority=[0] * len(PROMPTS)) |
| 76 | assert len(PROMPTS) == len(outputs) |
| 77 | |
| 78 | # Exception raised, if the length of priority does not match the length of prompts |
| 79 | with pytest.raises(VLLMValidationError): |
| 80 | outputs = llm.generate( |
| 81 | PROMPTS, sampling_params=None, priority=[0] * (len(PROMPTS) - 1) |
| 82 | ) |
| 83 | |
| 84 | # Exception raised, if the priority list is empty |
| 85 | with pytest.raises(VLLMValidationError): |
| 86 | outputs = llm.generate(PROMPTS, sampling_params=None, priority=[]) |
| 87 | |
| 88 | |
| 89 | def test_single_prompt_priority(llm: LLM): |
| 90 | # Single string prompts should be normalized to one request. |
| 91 | outputs = llm.generate(PROMPTS[0], sampling_params=None, priority=[0]) |
| 92 | assert len(outputs) == 1 |
| 93 | |
| 94 | |
| 95 | def test_max_model_len(vllm_runner): |
| 96 | max_model_len = 20 |
| 97 | with vllm_runner( |
| 98 | MODEL_NAME, |
| 99 | max_model_len=max_model_len, |
| 100 | gpu_memory_utilization=0.10, |
| 101 | enforce_eager=True, # reduce test time |
| 102 | ) as runner: |
| 103 | sampling_params = SamplingParams(max_tokens=max_model_len + 10) |
| 104 | outputs = runner.llm.generate(PROMPTS, sampling_params) |
| 105 | for output in outputs: |
| 106 | num_total_tokens = len(output.prompt_token_ids) + len( |
| 107 | output.outputs[0].token_ids |
| 108 | ) |
| 109 | # Total tokens must not exceed max_model_len. |
| 110 | # It can be less if generation finishes due to other reasons (e.g., EOS) |
| 111 | # before reaching the absolute model length limit. |
| 112 | assert num_total_tokens <= max_model_len |
| 113 | |
| 114 | |
| 115 | def test_log_stats(vllm_runner): |
| 116 | with vllm_runner( |
| 117 | MODEL_NAME, |
| 118 | disable_log_stats=False, |
| 119 | gpu_memory_utilization=0.10, |
| 120 | enforce_eager=True, # reduce test time |
| 121 | ) as runner: |
| 122 | outputs = runner.llm.generate(PROMPTS, sampling_params=None) |
| 123 | |
| 124 | # disable_log_stats is False, every output should have metrics |
| 125 | assert all(output.metrics is not None for output in outputs) |