|
| 1 | +# SPDX-License-Identifier: Apache-2.0 |
| 2 | +""" |
| 3 | +This example shows how to use vLLM for running offline inference |
| 4 | +with the correct prompt format on Qwen2.5-Omni (thinker only). |
| 5 | +""" |
| 6 | + |
| 7 | +from typing import NamedTuple |
| 8 | + |
| 9 | +import vllm.envs as envs |
| 10 | +from vllm import LLM, SamplingParams |
| 11 | +from vllm.assets.audio import AudioAsset |
| 12 | +from vllm.assets.image import ImageAsset |
| 13 | +from vllm.assets.video import VideoAsset |
| 14 | +from vllm.utils import FlexibleArgumentParser |
| 15 | + |
| 16 | + |
| 17 | +class QueryResult(NamedTuple): |
| 18 | + inputs: dict |
| 19 | + limit_mm_per_prompt: dict[str, int] |
| 20 | + |
| 21 | + |
| 22 | +# NOTE: The default `max_num_seqs` and `max_model_len` may result in OOM on |
| 23 | +# lower-end GPUs. |
| 24 | +# Unless specified, these settings have been tested to work on a single L4. |
| 25 | + |
| 26 | +default_system = ( |
| 27 | + "You are Qwen, a virtual human developed by the Qwen Team, Alibaba " |
| 28 | + "Group, capable of perceiving auditory and visual inputs, as well as " |
| 29 | + "generating text and speech.") |
| 30 | + |
| 31 | + |
| 32 | +def get_mixed_modalities_query() -> QueryResult: |
| 33 | + question = ("What is recited in the audio? " |
| 34 | + "What is the content of this image? Why is this video funny?") |
| 35 | + prompt = (f"<|im_start|>system\n{default_system}<|im_end|>\n" |
| 36 | + "<|im_start|>user\n<|audio_bos|><|AUDIO|><|audio_eos|>" |
| 37 | + "<|vision_bos|><|IMAGE|><|vision_eos|>" |
| 38 | + "<|vision_bos|><|VIDEO|><|vision_eos|>" |
| 39 | + f"{question}<|im_end|>\n" |
| 40 | + f"<|im_start|>assistant\n") |
| 41 | + return QueryResult( |
| 42 | + inputs={ |
| 43 | + "prompt": prompt, |
| 44 | + "multi_modal_data": { |
| 45 | + "audio": |
| 46 | + AudioAsset("mary_had_lamb").audio_and_sample_rate, |
| 47 | + "image": |
| 48 | + ImageAsset("cherry_blossom").pil_image.convert("RGB"), |
| 49 | + "video": |
| 50 | + VideoAsset(name="sample_demo_1.mp4", |
| 51 | + num_frames=16).np_ndarrays, |
| 52 | + }, |
| 53 | + }, |
| 54 | + limit_mm_per_prompt={ |
| 55 | + "audio": 1, |
| 56 | + "image": 1, |
| 57 | + "video": 1 |
| 58 | + }, |
| 59 | + ) |
| 60 | + |
| 61 | + |
| 62 | +def get_use_audio_in_video_query() -> QueryResult: |
| 63 | + question = ("Describe the content of the video, " |
| 64 | + "then convert what the baby say into text.") |
| 65 | + prompt = (f"<|im_start|>system\n{default_system}<|im_end|>\n" |
| 66 | + "<|im_start|>user\n<|vision_bos|><|VIDEO|><|vision_eos|>" |
| 67 | + f"{question}<|im_end|>\n" |
| 68 | + f"<|im_start|>assistant\n") |
| 69 | + asset = VideoAsset(name="sample_demo_1.mp4", num_frames=16) |
| 70 | + audio = asset.get_audio(sampling_rate=16000) |
| 71 | + assert not envs.VLLM_USE_V1, ("V1 does not support use_audio_in_video. " |
| 72 | + "Please launch this example with " |
| 73 | + "`VLLM_USE_V1=0`.") |
| 74 | + return QueryResult( |
| 75 | + inputs={ |
| 76 | + "prompt": prompt, |
| 77 | + "multi_modal_data": { |
| 78 | + "video": asset.np_ndarrays, |
| 79 | + "audio": audio, |
| 80 | + }, |
| 81 | + "mm_processor_kwargs": { |
| 82 | + "use_audio_in_video": True, |
| 83 | + }, |
| 84 | + }, |
| 85 | + limit_mm_per_prompt={ |
| 86 | + "audio": 1, |
| 87 | + "video": 1 |
| 88 | + }, |
| 89 | + ) |
| 90 | + |
| 91 | + |
| 92 | +def get_multi_audios_query() -> QueryResult: |
| 93 | + question = "Are these two audio clips the same?" |
| 94 | + prompt = (f"<|im_start|>system\n{default_system}<|im_end|>\n" |
| 95 | + "<|im_start|>user\n<|audio_bos|><|AUDIO|><|audio_eos|>" |
| 96 | + "<|audio_bos|><|AUDIO|><|audio_eos|>" |
| 97 | + f"{question}<|im_end|>\n" |
| 98 | + f"<|im_start|>assistant\n") |
| 99 | + return QueryResult( |
| 100 | + inputs={ |
| 101 | + "prompt": prompt, |
| 102 | + "multi_modal_data": { |
| 103 | + "audio": [ |
| 104 | + AudioAsset("winning_call").audio_and_sample_rate, |
| 105 | + AudioAsset("mary_had_lamb").audio_and_sample_rate, |
| 106 | + ], |
| 107 | + }, |
| 108 | + }, |
| 109 | + limit_mm_per_prompt={ |
| 110 | + "audio": 2, |
| 111 | + }, |
| 112 | + ) |
| 113 | + |
| 114 | + |
| 115 | +query_map = { |
| 116 | + "mixed_modalities": get_mixed_modalities_query, |
| 117 | + "use_audio_in_video": get_use_audio_in_video_query, |
| 118 | + "multi_audios": get_multi_audios_query, |
| 119 | +} |
| 120 | + |
| 121 | + |
| 122 | +def main(args): |
| 123 | + model_name = "Qwen/Qwen2.5-Omni-7B" |
| 124 | + query_result = query_map[args.query_type]() |
| 125 | + |
| 126 | + llm = LLM(model=model_name, |
| 127 | + max_model_len=5632, |
| 128 | + max_num_seqs=5, |
| 129 | + limit_mm_per_prompt=query_result.limit_mm_per_prompt, |
| 130 | + seed=args.seed) |
| 131 | + |
| 132 | + # We set temperature to 0.2 so that outputs can be different |
| 133 | + # even when all prompts are identical when running batch inference. |
| 134 | + sampling_params = SamplingParams(temperature=0.2, max_tokens=64) |
| 135 | + |
| 136 | + outputs = llm.generate(query_result.inputs, |
| 137 | + sampling_params=sampling_params) |
| 138 | + |
| 139 | + for o in outputs: |
| 140 | + generated_text = o.outputs[0].text |
| 141 | + print(generated_text) |
| 142 | + |
| 143 | + |
| 144 | +if __name__ == "__main__": |
| 145 | + parser = FlexibleArgumentParser( |
| 146 | + description='Demo on using vLLM for offline inference with ' |
| 147 | + 'audio language models') |
| 148 | + parser.add_argument('--query-type', |
| 149 | + '-q', |
| 150 | + type=str, |
| 151 | + default="mixed_modalities", |
| 152 | + choices=query_map.keys(), |
| 153 | + help='Query type.') |
| 154 | + parser.add_argument("--seed", |
| 155 | + type=int, |
| 156 | + default=None, |
| 157 | + help="Set the seed when initializing `vllm.LLM`.") |
| 158 | + |
| 159 | + args = parser.parse_args() |
| 160 | + main(args) |
0 commit comments