推理示例函数（仅生成文本，不生成语音） Args: model_path: 模型路径 audio_path: 输入音频路径 instruction: 输入音频指令，可选
(model_path, audio_path, instruction: Optional[str] = None)
| 30 | device = "cuda:0" if torch.cuda.is_available() else "cpu" |
| 31 | |
| 32 | def infer_example(model_path, audio_path, instruction: Optional[str] = None): |
| 33 | """ |
| 34 | 推理示例函数（仅生成文本，不生成语音） |
| 35 | |
| 36 | Args: |
| 37 | model_path: 模型路径 |
| 38 | audio_path: 输入音频路径 |
| 39 | instruction: 输入音频指令，可选 |
| 40 | """ |
| 41 | config = AutoConfig.from_pretrained(model_path) |
| 42 | processor = AutoProcessor.from_pretrained(model_path) |
| 43 | model = AutoModelForSeq2SeqLM.from_pretrained(model_path, config=config, torch_dtype=torch.bfloat16, device_map=device) |
| 44 | |
| 45 | # 生成参数 |
| 46 | model.sp_gen_kwargs.update({ |
| 47 | 'text_greedy': True, |
| 48 | 'disable_speech': True, |
| 49 | }) |
| 50 | |
| 51 | # 构建audio样例 |
| 52 | audio = [librosa.load(audio_path, sr=16000)[0]] |
| 53 | |
| 54 | if instruction is None: |
| 55 | conversation = [ |
| 56 | {"role": "system", "content": DEFAULT_S2T_PROMPT}, |
| 57 | {"role": "user", "content": AUDIO_TEMPLATE}, |
| 58 | ] |
| 59 | else: |
| 60 | conversation = [ |
| 61 | {"role": "system", "content": DEFAULT_S2T_PROMPT}, |
| 62 | {"role": "user", "content": AUDIO_TEMPLATE + "\n" + instruction}, |
| 63 | ] |
| 64 | |
| 65 | text = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False) |
| 66 | inputs = processor(text=text, audio=audio, return_tensors="pt", return_token_type_ids=False).to(model.device) |
| 67 | generate_ids, _ = model.generate(**inputs) |
| 68 | generate_ids = generate_ids[:, inputs.input_ids.size(1):] |
| 69 | generate_text = processor.decode(generate_ids[0], skip_special_tokens=True) |
| 70 | |
| 71 | print("generate_text: ", generate_text) |
| 72 | |
| 73 | |
| 74 | def infer_function_calling_example(model_path, audio_path): |