This default vicuna template formats inputs in the form of a chat history. For each message in the chat history: * the template will

Encoding thread: Take complete frames from audio_buffer, encode them into Opus, and place them into opus_bytes_queue

Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (

r""" input_features (`torch.float` of shape `(features_dim, N)`): N = sum(feature_lens) Float values of mel features extracted fro

r""" Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence toke

Get multimodal inputs for training.

Function Calling 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

多轮对话推理示例函数，这里我们采用assistant生成的text作为下一轮的输入 Args: model_path: 模型路径 audio_paths: 输入音频路径列表

Print the attention implementation being used. Args: config: The model configuration

Process messages and replace audio placeholders.

Async tasks for handling audio saving and inference

Saves the attributes of this processor (feature extractor, tokenizer...) in the specified directory so that it can be reloaded using

Async coroutine: Read encoded Opus data from the queue and send at fixed time intervals

Asynchronous coroutine: Read text from the text buffer queue and send it at 200ms intervals after audio transmission begins.

[Multiprocessing] Receiving output from the TTS process (only messages belonging to the current session)

[Multiprocessing] Send audio tokens to the TTS process

TTS independent process1, avoiding Python GIL Args: input_queue: (uuid, tokens_list, offset, finalize) output_queue: (uu