pip install transformers#  由于Qwen2.5-Omni的代码在Hugging Face transformers中目前处于未合并阶段，尚未并入主分支，官方建议从源代码构建：#  pip uninstall transformers#  pip install git+https://github.com/huggingface/transformers@3a1ead0aabed473eafe527915eea8c197d424356pip install accelerate# 安装qwen-omni相关依赖工具包pip install qwen-omni-utils[decord]

import soundfile as sf
from transformers import Qwen2_5OmniModel, Qwen2_5OmniProcessorfrom qwen_omni_utils import process_mm_info
# default: Load the model on the available device(s)model = Qwen2_5OmniModel.from_pretrained("Qwen/Qwen2.5-Omni-7B", torch_dtype="auto", device_map="auto")
# 我们建议启用 flash_attention_2 以获取更快的推理速度以及更低的显存占用.# model = Qwen2_5OmniModel.from_pretrained(#     "Qwen/Qwen2.5-Omni-7B",#     torch_dtype="auto",#     device_map="auto",#     attn_implementation="flash_attention_2",# )
processor = Qwen2_5OmniProcessor.from_pretrained("Qwen/Qwen2.5-Omni-7B")
conversation = [    {        "role": "system",        "content": "You are Qwen, a virtual human developed by the Qwen Team, Alibaba Group, capable of perceiving auditory and visual inputs, as well as generating text and speech.",    },    {        "role": "user",        "content": [            {"type": "video", "video": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2.5-Omni/draw.mp4"},        ],    },]
# Preparation for inferencetext = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False)audios, images, videos = process_mm_info(conversation, use_audio_in_video=True)inputs = processor(text=text, audios=audios, images=images, videos=videos, return_tensors="pt", padding=True)inputs = inputs.to(model.device).to(model.dtype)
# Inference: Generation of the output text and audiotext_ids, audio = model.generate(**inputs, use_audio_in_video=True)
text = processor.batch_decode(text_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)print(text)sf.write(    "output.wav",    audio.reshape(-1).detach().cpu().numpy(),    samplerate=24000,)

pip install git+https://github.com/huggingface/transformers@1d04f0d44251be5e236484f8c8a00e1c7aa69022pip install acceleratepip install qwen-omni-utilsgit clone -b qwen2_omni_public_v1 https://github.com/fyabc/vllm.gitcd vllmpip install .

import osimport torch
from transformers import Qwen2_5OmniProcessorfrom vllm import LLM, SamplingParamsfrom qwen_omni_utils import process_mm_info
# vLLM engine v1 not supported yetos.environ['VLLM_USE_V1'] = '0'
MODEL_PATH = "Qwen/Qwen2.5-Omni-7B"
llm = LLM(    model=MODEL_PATH, trust_remote_code=True, gpu_memory_utilization=0.9,    tensor_parallel_size=torch.cuda.device_count(),    limit_mm_per_prompt={'image': 1, 'video': 1, 'audio': 1},    seed=1234,)
sampling_params = SamplingParams(    temperature=1e-6,    max_tokens=512,)
processor = Qwen2_5OmniProcessor.from_pretrained(MODEL_PATH)
messages = [    {        "role": "system",        "content": "You are Qwen, a virtual human developed by the Qwen Team, Alibaba Group, capable of perceiving auditory and visual inputs, as well as generating text and speech.",    },    {        "role": "user",        "content": [            {"type": "video", "video": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2.5-Omni/draw.mp4"},        ],    },]
text = processor.apply_chat_template(    messages,    tokenize=False,    add_generation_prompt=True,)
audios, images, videos = process_mm_info(messages, use_audio_in_video=True)
inputs = {    'prompt': text[0],    'multi_modal_data': {},    "mm_processor_kwargs": {        "use_audio_in_video": True,    },}

if images is not None:    inputs['multi_modal_data']['image'] = imagesif videos is not None:    inputs['multi_modal_data']['video'] = videosif audios is not None:    inputs['multi_modal_data']['audio'] = audios
outputs = llm.generate(inputs, sampling_params=sampling_params)print(outputs[0].outputs[0].text)

git clone https://github.com/QwenLM/Qwen2.5-Omni.gitpip install -r requirements_web_demo.txt

python web_demo.py

python web_demo.py --flash-attn2