| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # Copyright 2025 The vLLM team. |
| 5 | # Copyright 2025 NVIDIA CORPORATION and the HuggingFace Inc. team. All rights |
| 6 | # reserved. |
| 7 | # |
| 8 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 9 | # you may not use this file except in compliance with the License. |
| 10 | # You may obtain a copy of the License at |
| 11 | # |
| 12 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 13 | # |
| 14 | # Unless required by applicable law or agreed to in writing, software |
| 15 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 16 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 17 | # See the License for the specific language governing permissions and |
| 18 | # limitations under the License. |
| 19 | |
| 20 | import json |
| 21 | import os |
| 22 | |
| 23 | import pytest |
| 24 | |
| 25 | from tests.models.registry import HF_EXAMPLE_MODELS |
| 26 | from vllm import LLM, SamplingParams |
| 27 | |
| 28 | MODEL_NAME = "nvidia/audio-flamingo-3-hf" |
| 29 | SINGLE_CONVERSATION = [ |
| 30 | { |
| 31 | "role": "user", |
| 32 | "content": [ |
| 33 | { |
| 34 | "type": "text", |
| 35 | "text": "What is surprising about the relationship between " |
| 36 | "the barking and the music?", |
| 37 | }, |
| 38 | { |
| 39 | "type": "audio_url", |
| 40 | "audio_url": { |
| 41 | "url": "https://huggingface.co/datasets/nvidia/AudioSkills/" |
| 42 | "resolve/main/assets/" |
| 43 | "dogs_barking_in_sync_with_the_music.wav", |
| 44 | }, |
| 45 | }, |
| 46 | ], |
| 47 | } |
| 48 | ] |
| 49 | BATCHED_CONVERSATIONS = [ |
| 50 | SINGLE_CONVERSATION, |
| 51 | [ |
| 52 | { |
| 53 | "role": "user", |
| 54 | "content": [ |
| 55 | { |
| 56 | "type": "text", |
| 57 | "text": "Why is the philosopher's name mentioned in the " |
| 58 | "lyrics? (A) To express a sense of nostalgia " |
| 59 | "(B) To indicate that language cannot express clearly, " |
| 60 | "satirizing the inversion of black and white in the world " |
| 61 | "(C) To add depth and complexity to the lyrics " |
| 62 | "(D) To showcase the wisdom and influence of the " |
| 63 | "philosopher", |
| 64 | }, |
| 65 | { |
| 66 | "type": "audio_url", |
| 67 | "audio_url": { |
| 68 | "url": "https://huggingface.co/datasets/nvidia/" |
| 69 | "AudioSkills/resolve/main/assets/" |
| 70 | "Ch6Ae9DT6Ko_00-04-03_00-04-31.wav", |
| 71 | }, |
| 72 | }, |
| 73 | ], |
| 74 | } |
| 75 | ], |
| 76 | ] |
| 77 | |
| 78 | |
| 79 | def get_fixture_path(filename): |
| 80 | return os.path.join( |
| 81 | os.path.dirname(__file__), "../../fixtures/audioflamingo3", filename |
| 82 | ) |
| 83 | |
| 84 | |
| 85 | def assert_output_matches(output, expected_text, expected_token_ids): |
| 86 | generated = output.outputs[0] |
| 87 | assert generated.text.strip() == expected_text |
| 88 | actual_token_ids = list(generated.token_ids) |
| 89 | assert ( |
| 90 | actual_token_ids == expected_token_ids |
| 91 | or actual_token_ids == expected_token_ids[:-1] |
| 92 | or actual_token_ids[:-1] == expected_token_ids |
| 93 | ) |
| 94 | |
| 95 | |
| 96 | @pytest.fixture(scope="module") |
| 97 | def llm(): |
| 98 | model_info = HF_EXAMPLE_MODELS.get_hf_info("AudioFlamingo3ForConditionalGeneration") |
| 99 | model_info.check_transformers_version(on_fail="skip") |
| 100 | |
| 101 | try: |
| 102 | return LLM( |
| 103 | model=MODEL_NAME, |
| 104 | dtype="bfloat16", |
| 105 | enforce_eager=True, |
| 106 | limit_mm_per_prompt={"audio": 1}, |
| 107 | ) |
| 108 | except Exception as e: |
| 109 | pytest.skip(f"Failed to load model {MODEL_NAME}: {e}") |
| 110 | |
| 111 | |
| 112 | def test_single_generation(llm): |
| 113 | fixture_path = get_fixture_path("expected_results_single.json") |
| 114 | if not os.path.exists(fixture_path): |
| 115 | pytest.skip(f"Fixture not found: {fixture_path}") |
| 116 | |
| 117 | with open(fixture_path) as f: |
| 118 | expected = json.load(f) |
| 119 | |
| 120 | sampling_params = SamplingParams(temperature=0.0, max_tokens=128) |
| 121 | |
| 122 | outputs = llm.chat( |
| 123 | messages=SINGLE_CONVERSATION, |
| 124 | sampling_params=sampling_params, |
| 125 | ) |
| 126 | assert_output_matches( |
| 127 | outputs[0], |
| 128 | expected["transcriptions"][0], |
| 129 | expected["token_ids"][0], |
| 130 | ) |
| 131 | |
| 132 | |
| 133 | def test_batched_generation(llm): |
| 134 | fixture_path = get_fixture_path("expected_results_batched.json") |
| 135 | if not os.path.exists(fixture_path): |
| 136 | pytest.skip(f"Fixture not found: {fixture_path}") |
| 137 | |
| 138 | with open(fixture_path) as f: |
| 139 | expected = json.load(f) |
| 140 | |
| 141 | sampling_params = SamplingParams(temperature=0.0, max_tokens=128) |
| 142 | |
| 143 | outputs = llm.chat( |
| 144 | messages=BATCHED_CONVERSATIONS, |
| 145 | sampling_params=sampling_params, |
| 146 | ) |
| 147 | |
| 148 | for i, output in enumerate(outputs): |
| 149 | assert_output_matches( |
| 150 | output, |
| 151 | expected["transcriptions"][i], |
| 152 | expected["token_ids"][i], |
| 153 | ) |
| 154 | |
| 155 | |
| 156 | def test_single_and_batched_generation_match(llm): |
| 157 | sampling_params = SamplingParams(temperature=0.0, max_tokens=128) |
| 158 | |
| 159 | single_output = llm.chat( |
| 160 | messages=SINGLE_CONVERSATION, |
| 161 | sampling_params=sampling_params, |
| 162 | )[0] |
| 163 | batched_output = llm.chat( |
| 164 | messages=BATCHED_CONVERSATIONS, |
| 165 | sampling_params=sampling_params, |
| 166 | )[0] |
| 167 | |
| 168 | assert single_output.outputs[0].text == batched_output.outputs[0].text |
| 169 | assert list(single_output.outputs[0].token_ids) == list( |
| 170 | batched_output.outputs[0].token_ids |
| 171 | ) |