init

LICENSE

@@ -0,0 +1,53 @@
+Tongyi Qianwen LICENSE AGREEMENT
+
+Tongyi Qianwen Release Date: August 3, 2023
+
+By clicking to agree or by using or distributing any portion or element of the Tongyi Qianwen Materials, you will be deemed to have recognized and accepted the content of this Agreement, which is effective immediately.
+
+1. Definitions
+    a. This Tongyi Qianwen LICENSE AGREEMENT (this "Agreement") shall mean the terms and conditions for use, reproduction, distribution and modification of the Materials as defined by this Agreement.
+    b. "We"(or "Us") shall mean Alibaba Cloud.
+    c. "You" (or "Your") shall mean a natural person or legal entity exercising the rights granted by this Agreement and/or using the Materials for any purpose and in any field of use.
+    d. "Third Parties" shall mean individuals or legal entities that are not under common control with Us or You.
+    e. "Tongyi Qianwen" shall mean the large language models (including Qwen model and Qwen-Chat model), and software and algorithms, consisting of trained model weights, parameters (including optimizer states), machine-learning model code, inference-enabling code, training-enabling code, fine-tuning enabling code and other elements of the foregoing distributed by Us.
+    f. "Materials" shall mean, collectively, Alibaba Cloud's proprietary Tongyi Qianwen and Documentation (and any portion thereof) made available under this Agreement.
+    g. "Source" form shall mean the preferred form for making modifications, including but not limited to model source code, documentation source, and configuration files.
+    h. "Object" form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, generated documentation,
+ and conversions to other media types.
+
+2. Grant of Rights
+You are granted a non-exclusive, worldwide, non-transferable and royalty-free limited license under Alibaba Cloud's intellectual property or other rights owned by Us embodied in the Materials to use, reproduce, distribute, copy, create derivative works of, and make modifications to the Materials.
+
+3. Redistribution
+You may reproduce and distribute copies of the Materials or derivative works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions:
+    a. You shall give any other recipients of the Materials or derivative works a copy of this Agreement;
+    b. You shall cause any modified files to carry prominent notices stating that You changed the files;
+    c. You shall retain in all copies of the Materials that You distribute the following attribution notices within a "Notice" text file distributed as a part of such copies: "Tongyi Qianwen is licensed under the Tongyi Qianwen LICENSE AGREEMENT, Copyright (c) Alibaba Cloud. All Rights Reserved."; and
+    d. You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such derivative works as a whole, provided Your use, reproduction, and distribution of the work otherwise complies with the terms and conditions of this Agreement.
+
+4. Restrictions
+If you are commercially using the Materials, and your product or service has more than 100 million monthly active users, You shall request a license from Us. You cannot exercise your rights under this Agreement without our express authorization.
+
+5. Rules of use
+    a. The Materials may be subject to export controls or restrictions in China, the United States or other countries or regions. You shall comply with applicable laws and regulations in your use of the Materials.
+    b. You can not use the Materials or any output therefrom to improve any other large language model (excluding Tongyi Qianwen or derivative works thereof).
+
+6. Intellectual Property
+    a. We retain ownership of all intellectual property rights in and to the Materials and derivatives made by or for Us. Conditioned upon compliance with the terms and conditions of this Agreement, with respect to any derivative works and modifications of the Materials that are made by you, you are and will be the owner of such derivative works and modifications.
+    b. No trademark license is granted to use the trade names, trademarks, service marks, or product names of Us, except as required to fulfill notice requirements under this Agreement or as required for reasonable and customary use in describing and redistributing the Materials.
+    c. If you commence a lawsuit or other proceedings (including a cross-claim or counterclaim in a lawsuit) against Us or any entity alleging that the Materials or any output therefrom, or any part of the foregoing, infringe any intellectual property or other right owned or licensable by you, then all licences granted to you under this Agreement shall terminate as of the date such lawsuit or other proceeding is commenced or brought.
+
+7. Disclaimer of Warranty and Limitation of Liability
+
+    a. We are not obligated to support, update, provide training for, or develop any further version of the Tongyi Qianwen Materials or to grant any license thereto.
+    b. THE MATERIALS ARE PROVIDED "AS IS" WITHOUT ANY EXPRESS OR IMPLIED WARRANTY OF ANY KIND INCLUDING WARRANTIES OF MERCHANTABILITY, NONINFRINGEMENT, OR FITNESS FOR A PARTICULAR PURPOSE. WE MAKE NO WARRANTY AND ASSUME NO RESPONSIBILITY FOR THE SAFETY OR STABILITY OF THE MATERIALS AND ANY OUTPUT THEREFROM.
+    c. IN NO EVENT SHALL WE BE LIABLE TO YOU FOR ANY DAMAGES, INCLUDING, BUT NOT LIMITED TO ANY DIRECT, OR INDIRECT, SPECIAL OR CONSEQUENTIAL DAMAGES ARISING FROM YOUR USE OR INABILITY TO USE THE MATERIALS OR ANY OUTPUT OF IT, NO MATTER HOW IT’S CAUSED.
+    d. You will defend, indemnify and hold harmless Us from and against any claim by any third party arising out of or related to your use or distribution of the Materials.
+
+8. Survival and Termination.
+    a. The term of this Agreement shall commence upon your acceptance of this Agreement or access to the Materials and will continue in full force and effect until terminated in accordance with the terms and conditions herein.
+    b. We may terminate this Agreement if you breach any of the terms or conditions of this Agreement. Upon termination of this Agreement, you must delete and cease use of the Materials. Sections 7 and 9 shall survive the termination of this Agreement.
+
+9. Governing Law and Jurisdiction.
+    a. This Agreement and any dispute arising out of or relating to it will be governed by the laws of China, without regard to conflict of law principles, and the UN Convention on Contracts for the International Sale of Goods does not apply to this Agreement.
+    b. The People's Courts in Hangzhou City shall have exclusive jurisdiction over any dispute arising out of this Agreement.

README.md

@@ -0,0 +1,136 @@
+---
+language:
+- zh
+- en
+tags:
+- qwen
+- multimodal
+- 音频理解
+license: other
+pipeline_tag: text-generation
+inference: false
+---
+
+# Qwen-Audio-Chat
+
+<br>
+
+<p align="center">
+    <img src="https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-Audio/audio_logo.jpg" width="400"/>
+<p>
+<br>
+
+<p align="center">
+        Qwen-Audio <a href="https://www.modelscope.cn/models/qwen/QWen-Audio/summary">🤖 <a> | <a href="https://huggingface.co/Qwen/Qwen-Audio">🤗</a>&nbsp ｜ Qwen-Audio-Chat <a href="https://www.modelscope.cn/models/qwen/QWen-Audio-Chat/summary">🤖 <a>| <a href="https://huggingface.co/Qwen/Qwen-Audio-Chat">🤗</a>&nbsp
+<br>
+&nbsp&nbsp<a href="https://modelscope.cn/studios/qwen/Qwen-Audio-Chat-Demo/summary">Demo</a>&nbsp | &nbsp&nbsp<a href="https://qwen-audio.github.io/Qwen-Audio/">Homepage</a>&nbsp ｜ &nbsp<a href="http://arxiv.org/abs/2311.07919">Paper</a>
+</p>
+<br><br>
+
+**Qwen-Audio** (Qwen Large Audio Language Model) is the multimodal version of the large model series, Qwen (abbr. Tongyi Qianwen), proposed by Alibaba Cloud. Qwen-Audio accepts diverse audio (human speech, natural sound, music and song) and text as inputs, outputs text. The contribution of Qwen-Audio include:
+
+- **Fundamental audio models**: Qwen-Audio is a fundamental multi-task audio-language model that supports various tasks, languages, and audio types, serving as a universal audio understanding model. Building upon Qwen-Audio, we develop Qwen-Audio-Chat through instruction fine-tuning, enabling multi-turn dialogues and supporting diverse audio-oriented scenarios.
+- **Multi-task learning framework for all types of audios**: To scale up audio-language pre-training, we address the challenge of variation in textual labels associated with different datasets by proposing a multi-task training framework, enabling knowledge sharing and avoiding one-to-many interference. Our model incorporates more than 30 tasks and extensive experiments show the model achieves strong performance.
+- **Strong Performance**: Experimental results show that Qwen-Audio achieves impressive performance across diverse benchmark tasks without requiring any task-specific fine-tuning, surpassing its counterparts. Specifically, Qwen-Audio achieves state-of-the-art results on the test set of Aishell1, cochlscene, ClothoAQA, and VocalSound.
+- **Flexible multi-run chat from audio and text input**: Qwen-Audio supports multiple-audio analysis, sound understading and reasoning, music appreciation, and tool usage for speech editing.
+
+**Qwen-Audio** 是阿里云研发的大规模音频语言模型（Large Audio Language Model）。Qwen-Audio 可以以多种音频 (包括说话人语音、自然音、音乐、歌声）和文本作为输入，并以文本作为输出。Qwen-Audio 系列模型的特点包括：
+
+- **音频基石模型**：Qwen-Audio是一个性能卓越的通用的音频理解模型，支持各种任务、语言和音频类型。在Qwen-Audio的基础上，我们通过指令微调开发了Qwen-Audio-Chat，支持多轮、多语言、多语言对话。Qwen-Audio和Qwen-Audio-Chat模型均已开源。
+- **兼容多种复杂音频的多任务学习框架**：为了避免由于数据收集来源不同以及任务类型不同，带来的音频到文本的一对多的干扰问题，我们提出了一种多任务训练框架，实现相似任务的知识共享，并尽可能减少不同任务之间的干扰。通过提出的框架，Qwen-Audio可以容纳训练超过30多种不同的音频任务；
+- **出色的性能**：Qwen-Audio在不需要任何任务特定的微调的情况下，在各种基准任务上取得了领先的结果。具体得，Qwen-Audio在Aishell1、cochlscene、ClothoAQA和VocalSound的测试集上都达到了SOTA；
+- **支持多轮音频和文本对话，支持各种语音场景**：Qwen-Audio-Chat支持声音理解和推理、音乐欣赏、多音频分析、多轮音频-文本交错对话以及外部语音工具的使用(如语音编辑)。
+
+
+We release Qwen-Audio and Qwen-Audio-Chat, which are pretrained model and Chat model respectively. For more details about Qwen-Audio, please refer to our [Github Repo](https://github.com/QwenLM/Qwen-Audio/tree/main). This repo is the one for Qwen-Audio-Chat.
+<br>
+
+目前，我们提供了Qwen-Audio和Qwen-Audio-Chat两个模型，分别为预训练模型和Chat模型。如果想了解更多关于信息，请点击[链接](https://github.com/QwenLM/Qwen-Audio/tree/main)查看Github仓库。本仓库为Qwen-Audio-Chat仓库。
+
+
+## Requirements
+* python 3.8 and above
+* pytorch 1.12 and above, 2.0 and above are recommended
+* CUDA 11.4 and above are recommended (this is for GPU users)
+* FFmpeg
+  <br>
+
+## Quickstart
+Below, we provide simple examples to show how to use Qwen-Audio with 🤗 Transformers.
+
+Before running the code, make sure you have setup the environment and installed the required packages. Make sure you meet the above requirements, and then install the dependent libraries.
+
+```bash
+pip install -r requirements.txt
+```
+Now you can start with Transformers. For more usage, please refer to [tutorial](https://github.com/QwenLM/Qwen-Audio/blob/main/TUTORIAL.md).
+
+#### 🤖 ModelScope
+
+To use Qwen-Audio for the inference, all you need to do is to input a few lines of codes as demonstrated below. However, **please make sure that you are using the latest code.**
+
+```python
+from modelscope import (
+    snapshot_download, AutoModelForCausalLM, AutoTokenizer, GenerationConfig
+)
+import torch
+model_id = 'qwen/Qwen-Audio-Chat'
+revision = 'master'
+
+model_dir = snapshot_download(model_id, revision=revision)
+torch.manual_seed(1234)
+
+tokenizer = AutoTokenizer.from_pretrained(model_dir, trust_remote_code=True)
+if not hasattr(tokenizer, 'model_dir'):
+    tokenizer.model_dir = model_dir
+
+# Note: The default behavior now has injection attack prevention off.
+tokenizer = AutoTokenizer.from_pretrained(model_dir, trust_remote_code=True)
+
+# use bf16
+# model = AutoModelForCausalLM.from_pretrained(model_dir, device_map="auto", trust_remote_code=True, bf16=True).eval()
+# use fp16
+# model = AutoModelForCausalLM.from_pretrained(model_dir, device_map="auto", trust_remote_code=True, fp16=True).eval()
+# use cpu only
+# model = AutoModelForCausalLM.from_pretrained(model_dir, device_map="cpu", trust_remote_code=True).eval()
+# use cuda device
+model = AutoModelForCausalLM.from_pretrained(model_dir, device_map="cuda", trust_remote_code=True).eval()
+
+
+# 1st dialogue turn
+query = tokenizer.from_list_format([
+    {'audio': 'https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-Audio/1272-128104-0000.flac'}, # Either a local path or an url
+    {'text': 'what does the person say?'},
+])
+response, history = model.chat(tokenizer, query=query, history=None)
+print(response)
+# The person says: "mister quilter is the apostle of the middle classes and we are glad to welcome his gospel".
+
+# 2nd dialogue turn
+response, history = model.chat(tokenizer, 'Find the start time and end time of the word "middle classes"', history=history)
+print(response)
+# The word "middle classes" starts at <|2.33|> seconds and ends at <|3.26|> seconds.
+```
+
+
+## License Agreement
+Researchers and developers are free to use the codes and model weights of Qwen-Audio-Chat. We also allow its commercial use. Check our license at [LICENSE](LICENSE) for more details.
+<br>
+
+## Citation
+If you find our paper and code useful in your research, please consider giving a star and citation
+
+```BibTeX
+@article{Qwen-Audio,
+  title={Qwen-Audio: Advancing Universal Audio Understanding via Unified Large-Scale Audio-Language Models},
+  author={Chu, Yunfei and Xu, Jin and Zhou, Xiaohuan and Yang, Qian and Zhang, Shiliang and Yan, Zhijie  and Zhou, Chang and Zhou, Jingren},
+  journal={arXiv preprint arXiv:2311.07919},
+  year={2023}
+}
+```
+<br>
+
+## Contact Us
+
+If you are interested to leave a message to either our research team or product team, feel free to send an email to qianwen_opensource@alibabacloud.com.
+

audio.py

@@ -0,0 +1,443 @@
+import base64
+import gzip
+from dataclasses import dataclass
+from typing import Dict, Iterable, Optional, List
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import Tensor, nn
+from subprocess import CalledProcessError, run, Popen, PIPE
+
+import os
+from functools import lru_cache
+from typing import Optional, Union
+
+def exact_div(x, y):
+    assert x % y == 0
+    return x // y
+
+# hard-coded audio hyperparameters
+SAMPLE_RATE = 16000
+N_FFT = 400
+N_MELS = 80
+HOP_LENGTH = 160
+CHUNK_LENGTH = 30
+N_SAMPLES = CHUNK_LENGTH * SAMPLE_RATE  # 480000 samples in a 30-second chunk
+N_FRAMES = exact_div(N_SAMPLES, HOP_LENGTH)  # 3000 frames in a mel spectrogram input
+
+N_SAMPLES_PER_TOKEN = HOP_LENGTH * 2  # the initial convolutions has stride 2
+FRAMES_PER_SECOND = exact_div(SAMPLE_RATE, HOP_LENGTH)  # 10ms per audio frame
+TOKENS_PER_SECOND = exact_div(SAMPLE_RATE, N_SAMPLES_PER_TOKEN)  # 20ms per audio token
+
+
+
+def get_T_after_cnn(L_in, dilation=1):
+    for (padding, kernel_size, stride) in eval("[(1,3,1)] + [(1,3,2)] "):
+        L_out = L_in + 2 * padding - dilation * (kernel_size - 1) - 1
+        L_out = 1 + L_out // stride
+        L_in = L_out
+    return L_out
+
+def load_bytesio_audio(content, sr: int = SAMPLE_RATE):
+    cmd = [
+        "ffmpeg",
+        "-nostdin",
+        "-threads", "0",
+        "-i", "pipe:",
+        "-f", "s16le",
+        "-ac", "1",
+        "-acodec", "pcm_s16le",
+        "-ar", str(sr),
+        "pipe:"
+    ]
+    p = Popen(cmd, stdin=PIPE, stdout=PIPE, stderr=PIPE, bufsize=-1)
+    out, _ = p.communicate(input=content)
+    return np.frombuffer(out, np.int16).flatten().astype(np.float32) / 32768.0
+
+def load_audio(file: str, sr: int = SAMPLE_RATE):
+    """
+    Open an audio file and read as mono waveform, resampling as necessary
+
+    Parameters
+    ----------
+    file: str
+        The audio file to open
+
+    sr: int
+        The sample rate to resample the audio if necessary
+
+    Returns
+    -------
+    A NumPy array containing the audio waveform, in float32 dtype.
+    """
+
+    # This launches a subprocess to decode audio while down-mixing
+    # and resampling as necessary.  Requires the ffmpeg CLI in PATH.
+    # fmt: off
+    cmd = [
+        "ffmpeg",
+        "-nostdin",
+        "-threads", "0",
+        "-i", file,
+        "-f", "s16le",
+        "-ac", "1",
+        "-acodec", "pcm_s16le",
+        "-ar", str(sr),
+        "-"
+    ]
+    # fmt: on
+    try:
+        out = run(cmd, capture_output=True, check=True).stdout
+    except CalledProcessError as e:
+        raise RuntimeError(f"Failed to load audio: {e.stderr.decode()}") from e
+
+    return np.frombuffer(out, np.int16).flatten().astype(np.float32) / 32768.0
+
+
+def pad_or_trim(array, length: int = N_SAMPLES, *, axis: int = -1):
+    """
+    Pad or trim the audio array to N_SAMPLES, as expected by the encoder.
+    """
+    if torch.is_tensor(array):
+        if array.shape[axis] > length:
+            array = array.index_select(
+                dim=axis, index=torch.arange(length, device=array.device)
+            )
+
+        if array.shape[axis] < length:
+            pad_widths = [(0, 0)] * array.ndim
+            pad_widths[axis] = (0, length - array.shape[axis])
+            array = F.pad(array, [pad for sizes in pad_widths[::-1] for pad in sizes])
+    else:
+        if array.shape[axis] > length:
+            array = array.take(indices=range(length), axis=axis)
+
+        if array.shape[axis] < length:
+            pad_widths = [(0, 0)] * array.ndim
+            pad_widths[axis] = (0, length - array.shape[axis])
+            array = np.pad(array, pad_widths)
+
+    return array
+
+def trim(array, length: int = N_SAMPLES, *, axis: int = -1):
+    """
+    Pad or trim the audio array to N_SAMPLES, as expected by the encoder.
+    """
+    if torch.is_tensor(array):
+        if array.shape[axis] > length:
+            array = array.index_select(
+                dim=axis, index=torch.arange(length, device=array.device)
+            )
+    else:
+        if array.shape[axis] > length:
+            array = array.take(indices=range(length), axis=axis)
+    return array
+
+
+@lru_cache(maxsize=None)
+def mel_filters(device, n_mels: int = N_MELS) -> torch.Tensor:
+    """
+    load the mel filterbank matrix for projecting STFT into a Mel spectrogram.
+    Allows decoupling librosa dependency; saved using:
+
+        np.savez_compressed(
+            "mel_filters.npz",
+            mel_80=librosa.filters.mel(sr=16000, n_fft=400, n_mels=80),
+        )
+    """
+    assert n_mels == 80, f"Unsupported n_mels: {n_mels}"
+    with np.load(
+        os.path.join(os.path.dirname(__file__), "mel_filters.npz") # todo
+        # os.path.join("assets", "mel_filters.npz")
+    ) as f:
+        return torch.from_numpy(f[f"mel_{n_mels}"]).to(device)
+
+
+def log_mel_spectrogram(
+    audio: Union[str, np.ndarray, torch.Tensor],
+    n_mels: int = N_MELS,
+    padding: int = 0,
+    device: Optional[Union[str, torch.device]] = None,
+):
+    """
+    Compute the log-Mel spectrogram of
+
+    Parameters
+    ----------
+    audio: Union[str, np.ndarray, torch.Tensor], shape = (*)
+        The path to audio or either a NumPy array or Tensor containing the audio waveform in 16 kHz
+
+    n_mels: int
+        The number of Mel-frequency filters, only 80 is supported
+
+    padding: int
+        Number of zero samples to pad to the right
+
+    device: Optional[Union[str, torch.device]]
+        If given, the audio tensor is moved to this device before STFT
+
+    Returns
+    -------
+    torch.Tensor, shape = (80, n_frames)
+        A Tensor that contains the Mel spectrogram
+    """
+    if not torch.is_tensor(audio):
+        if isinstance(audio, str):
+            audio = load_audio(audio)
+        audio = torch.from_numpy(audio)
+
+    if device is not None:
+        audio = audio.to(device)
+    if padding > 0:
+        audio = F.pad(audio, (0, padding))
+    window = torch.hann_window(N_FFT).to(audio.device)
+    stft = torch.stft(audio, N_FFT, HOP_LENGTH, window=window, return_complex=True)
+    magnitudes = stft[..., :-1].abs() ** 2
+
+    filters = mel_filters(audio.device, n_mels)
+    mel_spec = filters @ magnitudes
+
+    log_spec = torch.clamp(mel_spec, min=1e-10).log10()
+    log_spec = torch.maximum(log_spec, log_spec.max() - 8.0)
+    log_spec = (log_spec + 4.0) / 4.0
+    return log_spec
+
+
+@dataclass
+class ModelDimensions:
+    n_mels: int
+    n_audio_ctx: int
+    n_audio_state: int
+    n_audio_head: int
+    n_audio_layer: int
+    n_vocab: int
+    n_text_ctx: int
+    n_text_state: int
+    n_text_head: int
+    n_text_layer: int
+
+
+class LayerNorm(nn.LayerNorm):
+    def forward(self, x: Tensor) -> Tensor:
+        # return super().forward(x.float()).type(x.dtype)
+        return super().forward(x).type(x.dtype)
+
+
+
+
+class Linear(nn.Linear):
+    def forward(self, x: Tensor) -> Tensor:
+        return F.linear(
+            x,
+            self.weight.to(x.dtype),
+            None if self.bias is None else self.bias.to(x.dtype),
+        )
+
+
+class Conv1d(nn.Conv1d):
+    def _conv_forward(
+        self, x: Tensor, weight: Tensor, bias: Optional[Tensor]
+    ) -> Tensor:
+        return super()._conv_forward(
+            x, weight.to(x.dtype), None if bias is None else bias.to(x.dtype)
+        )
+
+
+def sinusoids(length, channels, max_timescale=10000):
+    """Returns sinusoids for positional embedding"""
+    assert channels % 2 == 0
+    log_timescale_increment = np.log(max_timescale) / (channels // 2 - 1)
+    inv_timescales = torch.exp(-log_timescale_increment * torch.arange(channels // 2))
+    scaled_time = torch.arange(length)[:, np.newaxis] * inv_timescales[np.newaxis, :]
+    return torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], dim=1)
+
+
+class MultiHeadAttention(nn.Module):
+    def __init__(self, n_state: int, n_head: int):
+        super().__init__()
+        self.n_head = n_head
+        self.query = Linear(n_state, n_state)
+        self.key = Linear(n_state, n_state, bias=False)
+        self.value = Linear(n_state, n_state)
+        self.out = Linear(n_state, n_state)
+
+    def forward(
+        self,
+        x: Tensor,
+        xa: Optional[Tensor] = None,
+        mask: Optional[Tensor] = None,
+        kv_cache: Optional[dict] = None,
+    ):
+        q = self.query(x)
+
+        if kv_cache is None or xa is None or self.key not in kv_cache:
+            # hooks, if installed (i.e. kv_cache is not None), will prepend the cached kv tensors;
+            # otherwise, perform key/value projections for self- or cross-attention as usual.
+            k = self.key(x if xa is None else xa)
+            v = self.value(x if xa is None else xa)
+        else:
+            # for cross-attention, calculate keys and values once and reuse in subsequent calls.
+            k = kv_cache[self.key]
+            v = kv_cache[self.value]
+
+        wv, qk = self.qkv_attention(q, k, v, mask)
+        return self.out(wv), qk
+
+    def qkv_attention(
+        self, q: Tensor, k: Tensor, v: Tensor, mask: Optional[Tensor] = None
+    ):
+        n_batch, n_ctx, n_state = q.shape
+        scale = (n_state // self.n_head) ** -0.25
+        q = q.view(*q.shape[:2], self.n_head, -1).permute(0, 2, 1, 3) * scale
+        k = k.view(*k.shape[:2], self.n_head, -1).permute(0, 2, 3, 1) * scale
+        v = v.view(*v.shape[:2], self.n_head, -1).permute(0, 2, 1, 3)
+
+        qk = q @ k
+        if mask is not None:
+            qk += mask
+
+        w = F.softmax(qk, dim=-1).to(q.dtype)
+        return (w @ v).permute(0, 2, 1, 3).flatten(start_dim=2), qk.detach()
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(self, n_state: int, n_head: int, cross_attention: bool = False):
+        super().__init__()
+
+        self.attn = MultiHeadAttention(n_state, n_head)
+        self.attn_ln = LayerNorm(n_state)
+
+        self.cross_attn = (
+            MultiHeadAttention(n_state, n_head) if cross_attention else None
+        )
+        self.cross_attn_ln = LayerNorm(n_state) if cross_attention else None
+
+        n_mlp = n_state * 4
+        self.mlp = nn.Sequential(
+            Linear(n_state, n_mlp), nn.GELU(), Linear(n_mlp, n_state)
+        )
+        self.mlp_ln = LayerNorm(n_state)
+
+    def forward(
+        self,
+        x: Tensor,
+        xa: Optional[Tensor] = None,
+        mask: Optional[Tensor] = None,
+        kv_cache: Optional[dict] = None,
+    ):
+        x = x + self.attn(self.attn_ln(x), mask=mask, kv_cache=kv_cache)[0]
+        if self.cross_attn:
+            x = x + self.cross_attn(self.cross_attn_ln(x), xa, kv_cache=kv_cache)[0]
+        x = x + self.mlp(self.mlp_ln(x))
+        return x
+
+
+class AudioEncoder(nn.Module):
+    def __init__(
+            self,
+            n_mels: int,
+            n_ctx: int,
+            n_state: int,
+            n_head: int,
+            n_layer: int,
+            output_dim: int = 512,
+            avg_pool: bool = True,
+            add_audio_bos_eos_token: bool = True,
+            **kwargs
+    ):
+        super().__init__()
+        self.conv1 = Conv1d(n_mels, n_state, kernel_size=3, padding=1)
+        self.conv2 = Conv1d(n_state, n_state, kernel_size=3, stride=2, padding=1)
+        self.register_buffer("positional_embedding", sinusoids(n_ctx, n_state))
+
+        self.blocks: Iterable[ResidualAttentionBlock] = nn.ModuleList(
+            [ResidualAttentionBlock(n_state, n_head) for _ in range(n_layer)]
+        )
+        self.ln_post = LayerNorm(n_state)
+
+        if avg_pool:
+            self.avg_pooler = nn.AvgPool1d(2, stride=2)
+        else:
+            self.avg_pooler = None
+        self.proj = nn.Linear(n_state, output_dim)
+        if add_audio_bos_eos_token:
+            self.audio_bos_eos_token = nn.Embedding(2, output_dim)
+        else:
+            self.audio_bos_eos_token = None
+        self.output_dim = output_dim
+        self.n_head = n_head
+
+    def forward(self, x: Tensor, padding_mask: Tensor=None, audio_lengths: Tensor=None):
+        """
+        x : torch.Tensor, shape = (batch_size, n_mels, n_ctx)
+            the mel spectrogram of the audio
+        """
+        x = x.to(dtype=self.conv1.weight.dtype,
+                 device=self.conv1.weight.device)
+        if audio_lengths is not None:
+            input_mel_len = audio_lengths[:,0] * 2
+            max_mel_len_in_batch = input_mel_len.max()
+            x = x[:, :, :max_mel_len_in_batch]
+        x = F.gelu(self.conv1(x))
+        x = F.gelu(self.conv2(x))
+        x = x.permute(0, 2, 1)  # B, L, D
+        bsz = x.size(0)
+        src_len = x.size(1)
+
+
+        self.input_positional_embedding = self.positional_embedding[:src_len]
+        assert x.shape[1:] == self.input_positional_embedding.shape, f"incorrect audio shape: {x.shape[1:], self.input_positional_embedding.shape}"
+        x = (x + self.input_positional_embedding).to(x.dtype)
+        if padding_mask is not None:
+            padding_mask = padding_mask.to(dtype=self.conv1.weight.dtype,
+                     device=self.conv1.weight.device)
+            batch_src_len = padding_mask.size(1)
+            x = x[:, :batch_src_len, :]
+            padding_mask = padding_mask.view(
+                bsz, -1, batch_src_len
+            )
+            padding_mask_ = padding_mask.all(1)
+            x[padding_mask_] = 0
+            key_padding_mask = padding_mask_.view(bsz, 1, 1, batch_src_len). \
+                expand(-1, self.n_head, -1, -1).reshape(bsz, self.n_head, 1, batch_src_len)
+            new_padding_mask = torch.zeros_like(key_padding_mask, dtype=x.dtype)
+            padding_mask = new_padding_mask.masked_fill(key_padding_mask, float("-inf"))
+
+        for block in self.blocks:
+            x = block(x, mask=padding_mask)
+
+
+        if self.avg_pooler:
+            x = x.permute(0, 2, 1)
+            x = self.avg_pooler(x)
+            x = x.permute(0, 2, 1)
+
+
+        x = self.ln_post(x)
+        x = self.proj(x)
+
+        if self.audio_bos_eos_token is not None:
+            bos = self.audio_bos_eos_token.weight[0][None, :]
+            eos = self.audio_bos_eos_token.weight[1][None, :]
+        else:
+            bos, eos = None, None
+        return x, bos, eos
+
+    def encode(self, input_audios: Tensor, input_audio_lengths: Tensor, audio_span_tokens: List):
+        real_input_audio_lens = input_audio_lengths[:, 0].tolist()
+        max_len_in_batch = max(real_input_audio_lens)
+        padding_mask = torch.ones([input_audios.size(0), max_len_in_batch]).to(dtype=self.conv1.weight.dtype,
+                                                                               device=self.conv1.weight.device)
+        for index in range(len(input_audios)):
+            padding_mask[index, :input_audio_lengths[index][0].item()] = 0
+        x, bos, eos = self(input_audios, padding_mask,input_audio_lengths)
+        output_audios = []
+        for i in range(len(audio_span_tokens)):
+            audio_span = audio_span_tokens[i]
+            audio = x[i][:audio_span-2]
+            if bos is not None:
+                audio = torch.concat([bos, audio, eos])
+            assert len(audio) == audio_span
+            output_audios.append(audio)
+        return output_audios

config.json

@@ -0,0 +1,60 @@
+{
+  "architectures": [
+    "QWenLMHeadModel"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_qwen.QWenConfig",
+    "AutoModelForCausalLM": "modeling_qwen.QWenLMHeadModel"
+  },
+    "audio": {
+    "add_audio_bos_eos_token": true,
+    "audio_start_id": 155163,
+    "avg_pool": true,
+    "n_ctx": 1500,
+    "n_head": 20,
+    "n_layer": 32,
+    "n_mels": 80,
+    "n_state": 1280,
+    "output_dim": 4096
+  },
+  "attn_dropout_prob": 0.0,
+  "bf16": false,
+  "emb_dropout_prob": 0.0,
+  "fp16": true,
+  "fp32": false,
+  "hidden_size": 4096,
+  "intermediate_size": 22016,
+  "initializer_range": 0.02,
+  "kv_channels": 128,
+  "layer_norm_epsilon": 1e-06,
+  "max_position_embeddings": 8192,
+  "model_type": "qwen",
+  "no_bias": true,
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "onnx_safe": null,
+  "quantization_config": {
+    "bits": 4,
+    "group_size": 128,
+    "damp_percent": 0.01,
+    "desc_act": false,
+    "static_groups": false,
+    "sym": true,
+    "true_sequential": true,
+    "model_name_or_path": null,
+    "model_file_base_name": "model",
+    "quant_method": "gptq"
+  },
+  "rotary_emb_base": 10000,
+  "rotary_pct": 1.0,
+  "scale_attn_weights": true,
+  "seq_length": 8192,
+  "tie_word_embeddings": false,
+  "tokenizer_class": "QWenTokenizer",
+  "transformers_version": "4.32.0",
+  "use_cache": true,
+  "use_dynamic_ntk": true,
+  "use_flash_attn": "auto",
+  "use_logn_attn": true,
+  "vocab_size": 155947
+}

configuration.json

@@ -0,0 +1,3 @@
+{"framework":"Pytorch",
+ "task":"multimodal-dialogue",
+ "allow_remote": true}

configuration_qwen.py

@@ -0,0 +1,71 @@
+# Copyright (c) Alibaba Cloud.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from transformers import PretrainedConfig
+
+
+class QWenConfig(PretrainedConfig):
+    model_type = "qwen"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=4096,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        emb_dropout_prob=0.0,
+        attn_dropout_prob=0.0,
+        layer_norm_epsilon=1e-6,
+        initializer_range=0.02,
+        max_position_embeddings=8192,
+        scale_attn_weights=True,
+        use_cache=True,
+        bf16=False,
+        fp16=False,
+        fp32=False,
+        kv_channels=128,
+        rotary_pct=1.0,
+        rotary_emb_base=10000,
+        use_dynamic_ntk=True,
+        use_logn_attn=True,
+        use_flash_attn="auto",
+        intermediate_size=22016,
+        no_bias=True,
+        tie_word_embeddings=False,
+        use_cache_quantization=False,
+        use_cache_kernel=False,
+        softmax_in_fp32=False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.emb_dropout_prob = emb_dropout_prob
+        self.attn_dropout_prob = attn_dropout_prob
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+        self.scale_attn_weights = scale_attn_weights
+        self.use_cache = use_cache
+        self.max_position_embeddings = max_position_embeddings
+        self.bf16 = bf16
+        self.fp16 = fp16
+        self.fp32 = fp32
+        self.kv_channels = kv_channels
+        self.rotary_pct = rotary_pct
+        self.rotary_emb_base = rotary_emb_base
+        self.use_dynamic_ntk = use_dynamic_ntk
+        self.use_logn_attn = use_logn_attn
+        self.use_flash_attn = use_flash_attn
+        self.no_bias = no_bias
+        self.use_cache_quantization = use_cache_quantization
+        self.use_cache_kernel = use_cache_kernel
+        self.softmax_in_fp32 = softmax_in_fp32
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs
+        )

cpp_kernels.py

@@ -0,0 +1,55 @@
+from torch.utils import cpp_extension
+import pathlib
+import os
+import subprocess
+
+def _get_cuda_bare_metal_version(cuda_dir):
+    raw_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "-V"],
+                                         universal_newlines=True)
+    output = raw_output.split()
+    release_idx = output.index("release") + 1
+    release = output[release_idx].split(".")
+    bare_metal_major = release[0]
+    bare_metal_minor = release[1][0]
+
+    return raw_output, bare_metal_major, bare_metal_minor
+
+def _create_build_dir(buildpath):
+    try:
+        os.mkdir(buildpath)
+    except OSError:
+        if not os.path.isdir(buildpath):
+            print(f"Creation of the build directory {buildpath} failed")
+
+# Check if cuda 11 is installed for compute capability 8.0
+cc_flag = []
+_, bare_metal_major, bare_metal_minor = _get_cuda_bare_metal_version(cpp_extension.CUDA_HOME)
+if int(bare_metal_major) >= 11:
+    cc_flag.append('-gencode')
+    cc_flag.append('arch=compute_80,code=sm_80')
+    if int(bare_metal_minor) >= 7:
+        cc_flag.append('-gencode')
+        cc_flag.append('arch=compute_90,code=sm_90')
+
+# Build path
+srcpath = pathlib.Path(__file__).parent.absolute()
+buildpath = srcpath / 'build'
+_create_build_dir(buildpath)
+
+def _cpp_extention_load_helper(name, sources, extra_cuda_flags):
+    return cpp_extension.load(
+        name=name,
+        sources=sources,
+        build_directory=buildpath,
+        extra_cflags=['-O3', ],
+        extra_cuda_cflags=['-O3',
+                           '-gencode', 'arch=compute_70,code=sm_70',
+                           '--use_fast_math'] + extra_cuda_flags + cc_flag,
+        verbose=1
+    )
+
+extra_flags = []
+
+cache_autogptq_cuda_256_sources = ["./cache_autogptq_cuda_256.cpp",
+           "./cache_autogptq_cuda_kernel_256.cu"]
+cache_autogptq_cuda_256 = _cpp_extention_load_helper("cache_autogptq_cuda_256", cache_autogptq_cuda_256_sources, extra_flags)

generation_config.json

@@ -0,0 +1,11 @@
+{
+  "chat_format": "chatml",
+  "eos_token_id": 151643,
+  "pad_token_id": 151643,
+  "max_window_size": 6144,
+  "max_new_tokens": 512,
+  "do_sample": true,
+  "top_k": 0,
+  "top_p": 0.5,
+  "transformers_version": "4.31.0"
+}

mel_filters.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd2cc75e70e36fcbdd8ffbc2499062f30094093e6bf2cbafa9859f59972b420b
+size 2048

modeling_qwen.py

@@ -0,0 +1,1425 @@
+# Copyright (c) Alibaba Cloud.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import copy
+import importlib
+import math
+import shutil
+import pathlib
+from typing import TYPE_CHECKING, Optional, Tuple, Union, Callable, List, Any, Generator, Dict
+import os
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+import warnings
+from torch.cuda.amp import autocast
+
+from torch.nn import CrossEntropyLoss
+from transformers import PreTrainedTokenizer, GenerationConfig, StoppingCriteriaList
+from transformers.generation.logits_process import LogitsProcessorList
+
+if TYPE_CHECKING:
+    from transformers.generation.streamers import BaseStreamer
+from transformers.generation.utils import GenerateOutput
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+
+try:
+    from einops import rearrange
+except ImportError:
+    rearrange = None
+from torch import nn
+
+SUPPORT_CUDA = torch.cuda.is_available()
+SUPPORT_BF16 = SUPPORT_CUDA and torch.cuda.is_bf16_supported()
+SUPPORT_FP16 = SUPPORT_CUDA and torch.cuda.get_device_capability(0)[0] >= 7
+SUPPORT_TORCH2 = hasattr(torch, '__version__') and int(torch.__version__.split(".")[0]) >= 2
+
+
+from .configuration_qwen import QWenConfig
+from .qwen_generation_utils import (
+    HistoryType,
+    make_context,
+    decode_tokens,
+    get_stop_words_ids,
+    StopWordsLogitsProcessor,
+)
+from .audio import AudioEncoder
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "qwen"
+_CONFIG_FOR_DOC = "QWenConfig"
+
+QWen_PRETRAINED_MODEL_ARCHIVE_LIST = ["qwen-7b"]
+
+_ERROR_BAD_CHAT_FORMAT = """\
+We detect you are probably using the pretrained model (rather than chat model) for chatting, since the chat_format in generation_config is not "chatml".
+If you are directly using the model downloaded from Huggingface, please make sure you are using our "Qwen/Qwen-7B-Chat" Huggingface model (rather than "Qwen/Qwen-7B") when you call model.chat().
+我们检测到您可能在使用预训练模型（而非chat模型）进行多轮chat，因为您当前在generation_config指定的chat_format，并未设置为我们在对话中所支持的"chatml"格式。
+如果您在直接使用我们从Huggingface提供的模型，请确保您在调用model.chat()时，使用的是"Qwen/Qwen-7B-Chat"模型（而非"Qwen/Qwen-7B"预训练模型）。
+"""
+
+_SENTINEL = object()
+_ERROR_STREAM_IN_CHAT = """\
+Pass argument `stream` to model.chat() is buggy, deprecated, and marked for removal. Please use model.chat_stream(...) instead of model.chat(..., stream=True).
+向model.chat()传入参数stream的用法可能存在Bug，该用法已被废弃，将在未来被移除。请使用model.chat_stream(...)代替model.chat(..., stream=True)。
+"""
+
+_ERROR_INPUT_CPU_QUERY_WITH_FLASH_ATTN_ACTIVATED = """\
+We detect you have activated flash attention support, but running model computation on CPU. Please make sure that your input data has been placed on GPU. If you actually want to run CPU computation, please following the readme and set device_map="cpu" to disable flash attention when loading the model (calling AutoModelForCausalLM.from_pretrained).
+检测到您的模型已激活了flash attention支持，但正在执行CPU运算任务。如使用flash attention，请您确认模型输入已经传到GPU上。如果您确认要执行CPU运算，请您在载入模型（调用AutoModelForCausalLM.from_pretrained）时，按照readme说法，指定device_map="cpu"以禁用flash attention。
+"""
+
+apply_rotary_emb_func = None
+rms_norm = None
+flash_attn_unpadded_func = None
+
+def _import_flash_attn():
+    global apply_rotary_emb_func, rms_norm, flash_attn_unpadded_func
+    try:
+        from flash_attn.layers.rotary import apply_rotary_emb_func as __apply_rotary_emb_func
+        apply_rotary_emb_func = __apply_rotary_emb_func
+    except ImportError:
+        logger.warn(
+            "Warning: import flash_attn rotary fail, please install FlashAttention rotary to get higher efficiency "
+            "https://github.com/Dao-AILab/flash-attention/tree/main/csrc/rotary"
+        )
+
+    try:
+        from flash_attn.ops.rms_norm import rms_norm as __rms_norm
+        rms_norm = __rms_norm
+    except ImportError:
+        logger.warn(
+            "Warning: import flash_attn rms_norm fail, please install FlashAttention layer_norm to get higher efficiency "
+            "https://github.com/Dao-AILab/flash-attention/tree/main/csrc/layer_norm"
+        )
+
+    try:
+        import flash_attn
+        if not hasattr(flash_attn, '__version__'):
+            from flash_attn.flash_attn_interface import flash_attn_unpadded_func as __flash_attn_unpadded_func
+        else:
+            if int(flash_attn.__version__.split(".")[0]) >= 2:
+                from flash_attn.flash_attn_interface import flash_attn_varlen_func as __flash_attn_unpadded_func
+            else:
+                from flash_attn.flash_attn_interface import flash_attn_unpadded_func as __flash_attn_unpadded_func
+        flash_attn_unpadded_func = __flash_attn_unpadded_func
+    except ImportError:
+        logger.warn(
+            "Warning: import flash_attn fail, please install FlashAttention to get higher efficiency "
+            "https://github.com/Dao-AILab/flash-attention"
+        )
+
+def quantize_cache_v(fdata, bits, qmax, qmin):
+    # b, s, head, h-dim->b, head, s, h-dim
+    qtype = torch.uint8
+    device = fdata.device
+    shape = fdata.shape
+
+    fdata_cal = torch.flatten(fdata, 2)
+    fmax = torch.amax(fdata_cal, dim=-1, keepdim=True)
+    fmin = torch.amin(fdata_cal, dim=-1, keepdim=True)
+    # Compute params
+    if qmax.device != fmax.device:
+        qmax = qmax.to(device)
+        qmin = qmin.to(device)
+    scale = (fmax - fmin) / (qmax - qmin)
+    zero = qmin - fmin / scale
+    scale = scale.unsqueeze(-1).repeat(1,1,shape[2],1).contiguous()
+    zero = zero.unsqueeze(-1).repeat(1,1,shape[2],1).contiguous()
+    # Quantize
+    res_data = fdata / scale + zero
+    qdata = torch.clamp(res_data, qmin, qmax).to(qtype)
+    return qdata.contiguous(), scale, zero
+
+def dequantize_cache_torch(qdata, scale, zero):
+    data = scale * (qdata - zero)
+    return data
+
+class FlashSelfAttention(torch.nn.Module):
+    def __init__(
+        self,
+        causal=False,
+        softmax_scale=None,
+        attention_dropout=0.0,
+    ):
+        super().__init__()
+        assert flash_attn_unpadded_func is not None, (
+            "Please install FlashAttention first, " "e.g., with pip install flash-attn"
+        )
+        assert (
+            rearrange is not None
+        ), "Please install einops first, e.g., with pip install einops"
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.dropout_p = attention_dropout
+
+    def unpad_input(self, hidden_states, attention_mask):
+        valid_mask = attention_mask.squeeze(1).squeeze(1).eq(0)
+        seqlens_in_batch = valid_mask.sum(dim=-1, dtype=torch.int32)
+        indices = torch.nonzero(valid_mask.flatten(), as_tuple=False).flatten()
+        max_seqlen_in_batch = seqlens_in_batch.max().item()
+        cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0))
+        hidden_states = hidden_states[indices]
+        return hidden_states, indices, cu_seqlens, max_seqlen_in_batch
+
+    def pad_input(self, hidden_states, indices, batch, seqlen):
+        output = torch.zeros(batch * seqlen, *hidden_states.shape[1:], device=hidden_states.device,
+                             dtype=hidden_states.dtype)
+        output[indices] = hidden_states
+        return rearrange(output, '(b s) ... -> b s ...', b=batch)
+
+    def forward(self, q, k, v, attention_mask=None):
+        assert all((i.dtype in [torch.float16, torch.bfloat16] for i in (q, k, v)))
+        assert all((i.is_cuda for i in (q, k, v)))
+        batch_size, seqlen_q = q.shape[0], q.shape[1]
+        seqlen_k = k.shape[1]
+        seqlen_out = seqlen_q
+
+        q, k, v = [rearrange(x, "b s ... -> (b s) ...") for x in [q, k, v]]
+        cu_seqlens_q = torch.arange(
+            0,
+            (batch_size + 1) * seqlen_q,
+            step=seqlen_q,
+            dtype=torch.int32,
+            device=q.device,
+        )
+
+        if batch_size > 1 and attention_mask is not None:
+            k, indices_k, cu_seqlens_k, seqlen_k = self.unpad_input(k, attention_mask)
+            if q.size(0) == v.size(0):
+                q = q[indices_k]
+                cu_seqlens_q = cu_seqlens_k
+                seqlen_q = seqlen_k
+            v = v[indices_k]
+        else:
+            cu_seqlens_k = torch.arange(
+                0,
+                (batch_size + 1) * seqlen_k,
+                step=seqlen_k,
+                dtype=torch.int32,
+                device=q.device,
+            )
+
+        if self.training:
+            assert seqlen_k == seqlen_q
+            is_causal = self.causal
+            dropout_p = self.dropout_p
+        else:
+            is_causal = seqlen_q == seqlen_k
+            dropout_p = 0
+
+        output = flash_attn_unpadded_func(
+            q,
+            k,
+            v,
+            cu_seqlens_q,
+            cu_seqlens_k,
+            seqlen_q,
+            seqlen_k,
+            dropout_p,
+            softmax_scale=self.softmax_scale,
+            causal=is_causal,
+        )
+        if batch_size > 1 and attention_mask is not None and seqlen_q == seqlen_k:
+            output = self.pad_input(output, indices_k, batch_size, seqlen_out)
+        else:
+            new_shape = (batch_size, output.shape[0] // batch_size) + output.shape[1:]
+            output = output.view(new_shape)
+        return output
+
+
+class QWenAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+
+        self.register_buffer("masked_bias", torch.tensor(-1e4), persistent=False)
+        self.seq_length = config.seq_length
+
+        self.hidden_size = config.hidden_size
+        self.split_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+
+        self.use_flash_attn = config.use_flash_attn
+        self.scale_attn_weights = True
+
+        self.projection_size = config.kv_channels * config.num_attention_heads
+
+        assert self.projection_size % config.num_attention_heads == 0
+        self.hidden_size_per_attention_head = (
+            self.projection_size // config.num_attention_heads
+        )
+
+        self.c_attn = nn.Linear(config.hidden_size, 3 * self.projection_size)
+
+        self.c_proj = nn.Linear(
+            config.hidden_size, self.projection_size, bias=not config.no_bias
+        )
+
+        self.is_fp32 = not (config.bf16 or config.fp16)
+        if (
+            self.use_flash_attn
+            and flash_attn_unpadded_func is not None
+            and not self.is_fp32
+        ):
+            self.core_attention_flash = FlashSelfAttention(
+                causal=True, attention_dropout=config.attn_dropout_prob
+            )
+        self.bf16 = config.bf16
+
+        self.use_dynamic_ntk = config.use_dynamic_ntk
+        self.use_logn_attn = config.use_logn_attn
+
+        logn_list = [
+            math.log(i, self.seq_length) if i > self.seq_length else 1
+            for i in range(1, 32768)
+        ]
+        logn_tensor = torch.tensor(logn_list)[None, :, None, None]
+        self.register_buffer("logn_tensor", logn_tensor, persistent=False)
+
+        self.attn_dropout = nn.Dropout(config.attn_dropout_prob)
+        self.softmax_in_fp32 = config.softmax_in_fp32 if hasattr(config, 'softmax_in_fp32') else False
+        self.use_cache_quantization = config.use_cache_quantization if hasattr(config, 'use_cache_quantization') else False
+        self.use_cache_kernel = config.use_cache_kernel if hasattr(config,'use_cache_kernel') else False
+        cache_dtype = torch.float
+        if self.bf16:
+            cache_dtype=torch.bfloat16
+        elif config.fp16:
+            cache_dtype = torch.float16
+        self.cache_qmax = torch.tensor(torch.iinfo(torch.uint8).max, dtype=cache_dtype)
+        self.cache_qmin = torch.tensor(torch.iinfo(torch.uint8).min, dtype=cache_dtype)
+
+        if config.use_cache_quantization and config.use_cache_kernel:
+            # pre check if the support files existing
+            module_root = pathlib.Path(__file__).parent
+            src_files = ("cache_autogptq_cuda_256.cpp", "cache_autogptq_cuda_kernel_256.cu")
+            if any(not (module_root/src).is_file() for src in src_files):
+                warnings.warn("KV cache kernel source files (.cpp and .cu) not found.")
+                self.cache_kernels = None
+            else:
+                try:
+                    from .cpp_kernels import cache_autogptq_cuda_256
+                    self.cache_kernels = cache_autogptq_cuda_256
+                except ImportError:
+                    warnings.warn("Failed to import KV cache kernels.")
+                    self.cache_kernels = None
+
+    def _attn(self, query, key, value, registered_causal_mask, attention_mask=None, head_mask=None):
+        device = query.device
+        if self.use_cache_quantization:
+            qk, qk_scale, qk_zero = key
+            if self.use_cache_kernel and self.cache_kernels is not None:
+                shape = query.shape[:-1] + (qk.shape[-2],)
+                attn_weights = torch.zeros(shape, dtype=torch.float16, device=device)
+                self.cache_kernels.vecquant8matmul_batched_faster_old(
+                    query.contiguous() if query.dtype == torch.float16 else query.to(torch.float16).contiguous(),
+                    qk.transpose(-1, -2).contiguous(),
+                    attn_weights,
+                    qk_scale.contiguous() if qk_scale.dtype == torch.float16 else qk_scale.to(torch.float16).contiguous(),
+                    qk_zero.contiguous()if qk_zero.dtype == torch.float16 else qk_zero.to(torch.float16).contiguous())
+                # attn_weights = attn_weights.to(query.dtype).contiguous()
+            else:
+                key = dequantize_cache_torch(qk, qk_scale, qk_zero)
+                attn_weights = torch.matmul(query, key.transpose(-1, -2))
+        else:
+            attn_weights = torch.matmul(query, key.transpose(-1, -2))
+
+        if self.scale_attn_weights:
+            if self.use_cache_quantization:
+                size_temp = value[0].size(-1)
+            else:
+                size_temp = value.size(-1)
+            attn_weights = attn_weights / torch.full(
+                [],
+                size_temp ** 0.5,
+                dtype=attn_weights.dtype,
+                device=attn_weights.device,
+            )
+        if self.use_cache_quantization:
+            query_length, key_length = query.size(-2), key[0].size(-2)
+        else:
+            query_length, key_length = query.size(-2), key.size(-2)
+        causal_mask = registered_causal_mask[
+            :, :, key_length - query_length : key_length, :key_length
+        ]
+        mask_value = torch.finfo(attn_weights.dtype).min
+        mask_value = torch.full([], mask_value, dtype=attn_weights.dtype).to(
+            attn_weights.device
+        )
+        attn_weights = torch.where(
+            causal_mask, attn_weights.to(attn_weights.dtype), mask_value
+        )
+
+        if attention_mask is not None:
+            attn_weights = attn_weights + attention_mask
+
+        if self.softmax_in_fp32:
+            attn_weights = nn.functional.softmax(attn_weights.float(), dim=-1)
+        else:
+            attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+
+        attn_weights = attn_weights.type(query.dtype)
+        attn_weights = self.attn_dropout(attn_weights)
+
+        if head_mask is not None:
+            attn_weights = attn_weights * head_mask
+
+        if self.use_cache_quantization:
+            qv, qv_scale, qv_zero = value
+            if self.use_cache_kernel and self.cache_kernels is not None:
+                shape = attn_weights.shape[:-1] + (query.shape[-1],)
+                attn_output = torch.zeros(shape, dtype=torch.float16, device=device)
+                self.cache_kernels.vecquant8matmul_batched_column_compression_faster_old(
+                    attn_weights.contiguous() if attn_weights.dtype == torch.float16 else attn_weights.to(torch.float16).contiguous(),
+                    qv.contiguous(),  # dtype: int32
+                    attn_output,
+                    qv_scale.contiguous() if qv_scale.dtype == torch.float16 else qv_scale.to(torch.float16).contiguous(),
+                    qv_zero.contiguous() if qv_zero.dtype == torch.float16 else qv_zero.to(torch.float16).contiguous())
+                if attn_output.dtype != query.dtype:
+                    attn_output = attn_output.to(query.dtype)
+                    attn_weights = attn_weights.to(query.dtype)
+            else:
+                value = dequantize_cache_torch(qv, qv_scale, qv_zero)
+                attn_output = torch.matmul(attn_weights, value)
+        else:
+            attn_output = torch.matmul(attn_weights, value)
+
+        attn_output = attn_output.transpose(1, 2)
+
+        return attn_output, attn_weights
+
+    def _split_heads(self, tensor, num_heads, attn_head_size):
+        new_shape = tensor.size()[:-1] + (num_heads, attn_head_size)
+        tensor = tensor.view(new_shape)
+        return tensor
+
+    def _merge_heads(self, tensor, num_heads, attn_head_size):
+        tensor = tensor.contiguous()
+        new_shape = tensor.size()[:-2] + (num_heads * attn_head_size,)
+        return tensor.view(new_shape)
+
+    def forward(
+        self,
+        hidden_states: Optional[Tuple[torch.FloatTensor]],
+        rotary_pos_emb_list: Optional[List[List[torch.Tensor]]] = None,
+        layer_past: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+    ):
+        mixed_x_layer = self.c_attn(hidden_states)
+
+        query, key, value = mixed_x_layer.split(self.split_size, dim=2)
+
+        query = self._split_heads(query, self.num_heads, self.head_dim)
+        key = self._split_heads(key, self.num_heads, self.head_dim)
+        value = self._split_heads(value, self.num_heads, self.head_dim)
+
+        if rotary_pos_emb_list is not None:
+            cur_len = query.shape[1]
+            if len(rotary_pos_emb_list) == 1:
+                rotary_pos_emb = rotary_pos_emb_list[0]
+                rotary_pos_emb = [i[:, -cur_len:, :, :] for i in rotary_pos_emb]
+                rotary_pos_emb = (rotary_pos_emb,) * 2
+                q_pos_emb, k_pos_emb = rotary_pos_emb
+                # Slice the pos emb for current inference
+                query = apply_rotary_pos_emb(query, q_pos_emb)
+                key = apply_rotary_pos_emb(key, k_pos_emb)
+            else:
+                query_list = []
+                key_list = []
+                for i, rotary_pos_emb in enumerate(rotary_pos_emb_list):
+                    rotary_pos_emb = [i[:, -cur_len:, :, :] for i in rotary_pos_emb]
+                    rotary_pos_emb = (rotary_pos_emb,) * 2
+                    q_pos_emb, k_pos_emb = rotary_pos_emb
+                    # Slice the pos emb for current inference
+                    query_list += [apply_rotary_pos_emb(query[i:i+1, :, :], q_pos_emb)]
+                    key_list += [apply_rotary_pos_emb(key[i:i+1, :, :], k_pos_emb)]
+                query = torch.cat(query_list, dim=0)
+                key = torch.cat(key_list, dim=0)
+
+        if self.use_cache_quantization:
+            key = quantize_cache_v(key.permute(0, 2, 1, 3),
+                                       bits=8,
+                                       qmin=self.cache_qmin,
+                                       qmax=self.cache_qmax)
+            value = quantize_cache_v(value.permute(0, 2, 1, 3),
+                                         bits=8,
+                                         qmin=self.cache_qmin,
+                                         qmax=self.cache_qmax)
+
+
+        if layer_past is not None:
+            past_key, past_value = layer_past[0], layer_past[1]
+            if self.use_cache_quantization:
+                # use_cache_quantization:
+                # present=((q_key,key_scale,key_zero_point),
+                #          (q_value,value_scale,value_zero_point))
+                key = (torch.cat((past_key[0], key[0]), dim=2),
+                       torch.cat((past_key[1], key[1]), dim=2),
+                       torch.cat((past_key[2], key[2]), dim=2))
+                value = (torch.cat((past_value[0], value[0]), dim=2),
+                         torch.cat((past_value[1], value[1]), dim=2),
+                         torch.cat((past_value[2], value[2]), dim=2))
+            else:
+                # not use_cache_quantization:
+                # present=(key,value)
+                key = torch.cat((past_key, key), dim=1)
+                value = torch.cat((past_value, value), dim=1)
+
+        if use_cache:
+            present = (key, value)
+        else:
+            present = None
+
+        if self.use_logn_attn and not self.training:
+            if self.use_cache_quantization:
+                seq_start = key[0].size(2) - query.size(1)
+                seq_end = key[0].size(2)
+            else:
+                seq_start = key.size(1) - query.size(1)
+                seq_end = key.size(1)
+            logn_tensor = self.logn_tensor[:, seq_start:seq_end, :, :].type_as(query)
+            query = query * logn_tensor.expand_as(query)
+
+        if (
+            self.use_flash_attn
+            and flash_attn_unpadded_func is not None
+            and not self.is_fp32
+            and query.is_cuda
+        ):
+            q, k, v = query, key, value
+            attn_output = self.core_attention_flash(q, k, v, attention_mask=attention_mask)
+        else:
+            registered_causal_mask = torch.tril(
+                torch.ones((key.size(1), key.size(1)), dtype=torch.bool, device=key.device)
+            ).view(1, 1, key.size(1), key.size(1))
+            query = query.permute(0, 2, 1, 3)
+            if not self.use_cache_quantization:
+                key = key.permute(0, 2, 1, 3)
+                value = value.permute(0, 2, 1, 3)
+            if (
+                registered_causal_mask is None
+                and self.use_flash_attn
+                and flash_attn_unpadded_func is not None
+                and not self.is_fp32
+                and not query.is_cuda
+            ):
+                raise Exception(_ERROR_INPUT_CPU_QUERY_WITH_FLASH_ATTN_ACTIVATED)
+
+            if not self.use_cache_quantization and SUPPORT_TORCH2:
+                causal_mask = registered_causal_mask[
+                    :, :, key.size(-2) - query.size(-2): key.size(-2), :key.size(-2)
+                ]
+                if attention_mask is not None:
+                    attention_mask = attention_mask.expand(
+                        -1, -1, causal_mask.size(2), -1
+                    ).masked_fill(~causal_mask, torch.finfo(query.dtype).min)
+                else:
+                    attention_mask = causal_mask
+                attn_output = F.scaled_dot_product_attention(
+                    query, key, value, attn_mask=attention_mask
+                ).transpose(1, 2)
+                attn_weight = None
+            else:
+                attn_output, attn_weight = self._attn(
+                    query, key, value, registered_causal_mask, attention_mask, head_mask
+                )
+        context_layer = self._merge_heads(
+            attn_output, self.num_heads, self.head_dim
+        )
+
+        attn_output = self.c_proj(context_layer)
+
+        outputs = (attn_output, present)
+        if output_attentions:
+            if (
+                self.use_flash_attn
+                and flash_attn_unpadded_func is not None
+                and not self.is_fp32
+            ):
+                raise ValueError("Cannot output attentions while using flash-attn")
+            else:
+                outputs += (attn_weight,)
+
+        return outputs
+
+
+class QWenMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.w1 = nn.Linear(
+            config.hidden_size, config.intermediate_size // 2, bias=not config.no_bias
+        )
+        self.w2 = nn.Linear(
+            config.hidden_size, config.intermediate_size // 2, bias=not config.no_bias
+        )
+        ff_dim_in = config.intermediate_size // 2
+        self.c_proj = nn.Linear(ff_dim_in, config.hidden_size, bias=not config.no_bias)
+
+    def forward(self, hidden_states):
+        a1 = self.w1(hidden_states)
+        a2 = self.w2(hidden_states)
+        intermediate_parallel = a1 * F.silu(a2)
+        output = self.c_proj(intermediate_parallel)
+        return output
+
+class QWenBlock(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        hidden_size = config.hidden_size
+        self.bf16 = config.bf16
+
+        self.ln_1 = RMSNorm(
+            hidden_size,
+            eps=config.layer_norm_epsilon,
+        )
+        self.attn = QWenAttention(config)
+        self.ln_2 = RMSNorm(
+            hidden_size,
+            eps=config.layer_norm_epsilon,
+        )
+
+        self.mlp = QWenMLP(config)
+
+    def forward(
+        self,
+        hidden_states: Optional[Tuple[torch.FloatTensor]],
+        rotary_pos_emb_list: Optional[List[List[torch.Tensor]]] = None,
+        layer_past: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+    ):
+        layernorm_output = self.ln_1(hidden_states)
+
+        attn_outputs = self.attn(
+            layernorm_output,
+            rotary_pos_emb_list,
+            layer_past=layer_past,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+        )
+        attn_output = attn_outputs[0]
+
+        outputs = attn_outputs[1:]
+
+        residual = hidden_states
+        layernorm_input = attn_output + residual
+
+        layernorm_output = self.ln_2(layernorm_input)
+
+        residual = layernorm_input
+        mlp_output = self.mlp(layernorm_output)
+        hidden_states = residual + mlp_output
+
+        if use_cache:
+            outputs = (hidden_states,) + outputs
+        else:
+            outputs = (hidden_states,) + outputs[1:]
+
+        return outputs
+
+
+class QWenPreTrainedModel(PreTrainedModel):
+    config_class = QWenConfig
+    base_model_prefix = "transformer"
+    is_parallelizable = False
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["QWenBlock"]
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, RMSNorm):
+            module.weight.data.fill_(1.0)
+
+        for name, p in module.named_parameters():
+            if name == "c_proj.weight":
+                p.data.normal_(
+                    mean=0.0,
+                    std=(
+                        self.config.initializer_range
+                        / math.sqrt(2 * self.config.num_hidden_layers)
+                    ),
+                )
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, QWenModel):
+            module.gradient_checkpointing = value
+
+
+class QWenModel(QWenPreTrainedModel):
+    _keys_to_ignore_on_load_missing = ["attn.masked_bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.vocab_size = config.vocab_size
+        self.num_hidden_layers = config.num_hidden_layers
+        self.embed_dim = config.hidden_size
+        self.use_cache_quantization = self.config.use_cache_quantization if hasattr(self.config, 'use_cache_quantization') else False
+
+        self.gradient_checkpointing = False
+        self.use_dynamic_ntk = config.use_dynamic_ntk
+        self.seq_length = config.seq_length
+
+        self.wte = nn.Embedding(self.vocab_size, self.embed_dim)
+
+        self.drop = nn.Dropout(config.emb_dropout_prob)
+
+        if config.rotary_pct == 1.0:
+            self.rotary_ndims = None
+        else:
+            assert config.rotary_pct < 1
+            self.rotary_ndims = int(
+                config.kv_channels * config.rotary_pct
+            )
+        dim = (
+            self.rotary_ndims
+            if self.rotary_ndims is not None
+            else config.kv_channels
+        )
+        self.rotary_emb = RotaryEmbedding(dim, base=config.rotary_emb_base)
+
+        self.use_flash_attn = config.use_flash_attn
+        self.is_fp32 = not (config.bf16 or config.fp16)
+
+        self.h = nn.ModuleList(
+            [
+                QWenBlock(
+                    config
+                )
+                for i in range(config.num_hidden_layers)
+            ]
+        )
+        self.ln_f = RMSNorm(
+            self.embed_dim,
+            eps=config.layer_norm_epsilon,
+        )
+
+        self.audio = AudioEncoder(**config.audio)
+
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.wte
+
+    def set_input_embeddings(self, new_embeddings):
+        self.wte = new_embeddings
+
+    def get_ntk_alpha(self, true_seq_len):
+        context_value = math.log(true_seq_len / self.seq_length, 2) + 1
+        ntk_alpha = 2 ** math.ceil(context_value) - 1
+        ntk_alpha = max(ntk_alpha, 1)
+        return ntk_alpha
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        audio_info: Dict = None
+    ):
+        if past_key_values is None and torch.any(input_ids == self.config.audio['audio_start_id']):
+            bos_pos = torch.where(input_ids == self.config.audio['audio_start_id'])
+            eos_pos = torch.where(input_ids == self.config.audio['audio_start_id'] + 1)
+            assert (bos_pos[0] == eos_pos[0]).all()
+            audio_pos = torch.stack((bos_pos[0], bos_pos[1], eos_pos[1]), dim=1)
+            if isinstance(audio_info, Dict):
+                audios = audio_info["input_audios"]
+                audio_span_tokens = audio_info["audio_span_tokens"]
+                input_audio_lengths = audio_info["input_audio_lengths"]
+                audios = self.audio.encode(audios,input_audio_lengths, audio_span_tokens)
+            else:
+                audios = torch.concat([_["input_audios"] for _ in audio_info])
+                input_audio_lengths = torch.concat([_["input_audio_lengths"] for _ in audio_info])
+                audio_span_tokens = []
+                for _ in audio_info:
+                    audio_span_tokens.extend(_['audio_span_tokens'])
+                audios = self.audio.encode(audios, input_audio_lengths, audio_span_tokens)
+
+        else:
+            audios = None
+
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError(
+                "You cannot specify both input_ids and inputs_embeds at the same time"
+            )
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+            batch_size = input_ids.shape[0]
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+            batch_size = inputs_embeds.shape[0]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if token_type_ids is not None:
+            token_type_ids = token_type_ids.view(-1, input_shape[-1])
+        if position_ids is not None:
+            position_ids = position_ids.view(-1, input_shape[-1])
+
+        if past_key_values is None:
+            past_length = 0
+            past_key_values = tuple([None] * len(self.h))
+        else:
+            if self.use_cache_quantization:
+                past_length = past_key_values[0][0][0].size(2)
+            else:
+                past_length = past_key_values[0][0].size(-2)
+        if position_ids is None:
+            position_ids = torch.arange(
+                past_length,
+                input_shape[-1] + past_length,
+                dtype=torch.long,
+                device=device,
+            )
+            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
+
+        if attention_mask is not None:
+            if batch_size <= 0:
+                raise ValueError("batch_size has to be defined and > 0")
+            attention_mask = attention_mask.view(batch_size, -1)
+            attention_mask = attention_mask[:, None, None, :]
+            attention_mask = attention_mask.to(dtype=self.dtype)
+            attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
+
+        encoder_attention_mask = None
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.wte(input_ids)
+        hidden_states = inputs_embeds
+
+        kv_seq_len = hidden_states.size()[1]
+        if past_key_values[0] is not None:
+            # past key values[0][0] shape: bs * seq_len * head_num * dim
+            if self.use_cache_quantization:
+                kv_seq_len += past_key_values[0][0][0].shape[2]
+            else:
+                kv_seq_len += past_key_values[0][0].shape[1]
+
+        if self.training or not self.use_dynamic_ntk:
+            ntk_alpha_list = [1.0]
+        elif kv_seq_len != hidden_states.size()[1]:
+            ntk_alpha_list = self.rotary_emb._ntk_alpha_cached_list
+        else:
+            ntk_alpha_list = []
+            if attention_mask is not None and kv_seq_len > self.seq_length:
+                true_seq_lens = attention_mask.squeeze(1).squeeze(1).eq(0).sum(dim=-1, dtype=torch.int32)
+                for i in range(hidden_states.size()[0]):
+                    true_seq_len = true_seq_lens[i].item()
+                    ntk_alpha = self.get_ntk_alpha(true_seq_len)
+                    ntk_alpha_list.append(ntk_alpha)
+            else:
+                ntk_alpha = self.get_ntk_alpha(kv_seq_len)
+                ntk_alpha_list.append(ntk_alpha)
+        self.rotary_emb._ntk_alpha_cached_list = ntk_alpha_list
+        rotary_pos_emb_list = [
+            self.rotary_emb(kv_seq_len, ntk_alpha=ntk_alpha) for ntk_alpha in ntk_alpha_list
+        ]
+
+        hidden_states = self.drop(hidden_states)
+        if audios is not None:
+            for idx, (i, a, b) in enumerate(audio_pos):
+                hidden_states[i][a : b+1] = audios[idx]
+        output_shape = input_shape + (hidden_states.size(-1),)
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        presents = () if use_cache else None
+        all_self_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, use_cache, output_attentions)
+
+                    return custom_forward
+
+                outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    rotary_pos_emb_list,
+                    None,
+                    attention_mask,
+                    head_mask[i],
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+
+                outputs = block(
+                    hidden_states,
+                    layer_past=layer_past,
+                    rotary_pos_emb_list=rotary_pos_emb_list,
+                    attention_mask=attention_mask,
+                    head_mask=head_mask[i],
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                )
+
+            hidden_states = outputs[0]
+            if use_cache is True:
+                presents = presents + (outputs[1],)
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
+
+        hidden_states = self.ln_f(hidden_states)
+        hidden_states = hidden_states.view(output_shape)
+        # Add last hidden state
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v for v in [hidden_states, presents, all_hidden_states] if v is not None
+            )
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+
+class QWenLMHeadModel(QWenPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.rotary_emb\.inv_freq"]
+    _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.masked_bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        assert (
+            config.bf16 + config.fp16 + config.fp32 <= 1
+        ), "Only one of \"bf16\", \"fp16\", \"fp32\" can be true"
+
+        autoset_precision = config.bf16 + config.fp16 + config.fp32 == 0
+
+        if autoset_precision:
+            if SUPPORT_BF16:
+                logger.warn(
+                    "The model is automatically converting to bf16 for faster inference. "
+                    "If you want to disable the automatic precision, please manually add bf16/fp16/fp32=True to \"AutoModelForCausalLM.from_pretrained\"."
+                )
+                config.bf16 = True
+            elif SUPPORT_FP16:
+                logger.warn(
+                    "The model is automatically converting to fp16 for faster inference. "
+                    "If you want to disable the automatic precision, please manually add bf16/fp16/fp32=True to \"AutoModelForCausalLM.from_pretrained\"."
+                )
+                config.fp16 = True
+            else:
+                config.fp32 = True
+
+        if config.bf16 and SUPPORT_CUDA and not SUPPORT_BF16:
+            logger.warn("Your device does NOT seem to support bf16, you can switch to fp16 or fp32 by by passing fp16/fp32=True in \"AutoModelForCausalLM.from_pretrained\".")
+        if config.fp16 and SUPPORT_CUDA and not SUPPORT_FP16:
+            logger.warn("Your device does NOT support faster inference with fp16, please switch to fp32 which is likely to be faster")
+        if config.fp32:
+            if SUPPORT_BF16:
+                logger.warn("Your device support faster inference by passing bf16=True in \"AutoModelForCausalLM.from_pretrained\".")
+            elif SUPPORT_FP16:
+                logger.warn("Your device support faster inference by passing fp16=True in \"AutoModelForCausalLM.from_pretrained\".")
+
+        if config.use_flash_attn == "auto":
+            if config.bf16 or config.fp16:
+                logger.warn("Try importing flash-attention for faster inference...")
+                config.use_flash_attn = True
+            else:
+                config.use_flash_attn = False
+        if config.use_flash_attn and config.fp32:
+            logger.warn("Flash attention will be disabled because it does NOT support fp32.")
+
+        if config.use_flash_attn:
+            _import_flash_attn()
+
+        self.transformer = QWenModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        if config.bf16:
+            self.transformer.bfloat16()
+            self.lm_head.bfloat16()
+        if config.fp16:
+            self.transformer.half()
+            self.lm_head.half()
+        self.post_init()
+
+    @classmethod
+    def from_pretrained(
+            cls,
+            pretrained_model_name_or_path: Optional[Union[str, os.PathLike]],
+            *model_args,
+            config=None,
+            cache_dir=None,
+            **kwargs,
+    ):
+        if os.path.isdir(pretrained_model_name_or_path):
+            # Local Directory of Models
+            mel_filters_path = os.path.join(pretrained_model_name_or_path, 'mel_filters.npz')
+            tgt_cache_path = os.path.join(os.path.dirname(__file__), 'mel_filters.npz')
+            shutil.copy(mel_filters_path, tgt_cache_path)
+        else:
+            # Loading from huggingface repo
+            from huggingface_hub import hf_hub_download
+            hf_hub_download(repo_id=pretrained_model_name_or_path, filename="mel_filters.npz",
+                            token=kwargs.get('token', None), local_dir=os.path.dirname(__file__))
+        return super().from_pretrained(pretrained_model_name_or_path, *model_args, config=config, cache_dir=cache_dir,
+                                       **kwargs)
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def prepare_inputs_for_generation(
+        self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs
+    ):
+        audio_info = kwargs.pop("audio_info", None)
+        token_type_ids = kwargs.get("token_type_ids", None)
+        if past_key_values:
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+            if token_type_ids is not None:
+                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
+
+        attention_mask = kwargs.get("attention_mask", None)
+        position_ids = kwargs.get("position_ids", None)
+
+        if attention_mask is not None and position_ids is None:
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
+        else:
+            position_ids = None
+
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "position_ids": position_ids,
+                "attention_mask": attention_mask,
+                "token_type_ids": token_type_ids,
+                "audio_info": audio_info
+            }
+        )
+        return model_inputs
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        audio_info: Dict = None
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            audio_info=audio_info
+        )
+        hidden_states = transformer_outputs[0]
+
+        lm_logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            labels = labels.to(lm_logits.device)
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(
+                shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
+            )
+
+        if not return_dict:
+            output = (lm_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    @staticmethod
+    def _reorder_cache(
+        past_key_values: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
+    ) -> Tuple[Tuple[torch.Tensor]]:
+
+        return tuple(
+            tuple(
+                past_state.index_select(0, beam_idx.to(past_state.device))
+                for past_state in layer_past
+            )
+            for layer_past in past_key_values
+        )
+
+    def chat(
+        self,
+        tokenizer: PreTrainedTokenizer,
+        query: str,
+        history: Optional[HistoryType],
+        system: str = "You are a helpful assistant.",
+        append_history: bool = True,
+        stream: Optional[bool] = _SENTINEL,
+        stop_words_ids: Optional[List[List[int]]] = None,
+        generation_config: Optional[GenerationConfig] = None,
+        **kwargs,
+    ) -> Tuple[str, HistoryType]:
+        generation_config = generation_config if generation_config is not None else self.generation_config
+
+        assert stream is _SENTINEL, _ERROR_STREAM_IN_CHAT
+        assert generation_config.chat_format == 'chatml', _ERROR_BAD_CHAT_FORMAT
+        if history is None:
+            history = []
+        else:
+            # make a copy of the user's input such that is is left untouched
+            history = copy.deepcopy(history)
+
+        if stop_words_ids is None:
+            stop_words_ids = []
+
+        max_window_size = kwargs.get('max_window_size', None)
+        if max_window_size is None:
+            max_window_size = generation_config.max_window_size
+        raw_text, context_tokens, audio_info = make_context(
+            tokenizer,
+            query,
+            history=history,
+            system=system,
+            max_window_size=max_window_size,
+            chat_format=generation_config.chat_format,
+        )
+
+        stop_words_ids.extend(get_stop_words_ids(
+            generation_config.chat_format, tokenizer
+        ))
+        input_ids = torch.tensor([context_tokens]).to(self.device)
+        kwargs['audio_info'] = audio_info
+        outputs = self.generate(
+                    input_ids,
+                    stop_words_ids=stop_words_ids,
+                    return_dict_in_generate=False,
+                    generation_config=generation_config,
+                    **kwargs,
+                )
+
+        response = decode_tokens(
+            outputs[0],
+            tokenizer,
+            raw_text_len=len(raw_text),
+            context_length=len(context_tokens),
+            chat_format=generation_config.chat_format,
+            verbose=False,
+            errors='replace',
+            audio_info=audio_info
+        )
+
+        # as history is a copy of the user inputs,
+        # we can always return the new turn to the user.
+        # separating input history and output history also enables the user
+        # to implement more complex history management
+        history.append((query, response))
+
+        return response, history
+
+    def chat_stream(
+            self,
+            tokenizer: PreTrainedTokenizer,
+            query: str,
+            history: Optional[HistoryType],
+            system: str = "You are a helpful assistant.",
+            stop_words_ids: Optional[List[List[int]]] = None,
+            logits_processor: Optional[LogitsProcessorList] = None,
+            generation_config: Optional[GenerationConfig] = None,
+            **kwargs,
+    ) -> Generator[str, Any, None]:
+        generation_config = generation_config if generation_config is not None else self.generation_config
+        assert generation_config.chat_format == 'chatml', _ERROR_BAD_CHAT_FORMAT
+        if history is None:
+            history = []
+        if stop_words_ids is None:
+            stop_words_ids = []
+
+        max_window_size = kwargs.get('max_window_size', None)
+        if max_window_size is None:
+            max_window_size = generation_config.max_window_size
+        raw_text, context_tokens = make_context(
+            tokenizer,
+            query,
+            history=history,
+            system=system,
+            max_window_size=max_window_size,
+            chat_format=generation_config.chat_format,
+        )
+
+        stop_words_ids.extend(get_stop_words_ids(
+            generation_config.chat_format, tokenizer
+        ))
+        if stop_words_ids is not None:
+            stop_words_logits_processor = StopWordsLogitsProcessor(
+                stop_words_ids=stop_words_ids,
+                eos_token_id=generation_config.eos_token_id,
+            )
+            if logits_processor is None:
+                logits_processor = LogitsProcessorList([stop_words_logits_processor])
+            else:
+                logits_processor.append(stop_words_logits_processor)
+        input_ids = torch.tensor([context_tokens]).to(self.device)
+
+        from transformers_stream_generator.main import NewGenerationMixin, StreamGenerationConfig
+        self.__class__.generate_stream = NewGenerationMixin.generate
+        self.__class__.sample_stream = NewGenerationMixin.sample_stream
+        stream_config = StreamGenerationConfig(**generation_config.to_dict(), do_stream=True)
+
+        def stream_generator():
+            outputs = []
+            for token in self.generate_stream(
+                    input_ids,
+                    return_dict_in_generate=False,
+                    generation_config=stream_config,
+                    logits_processor=logits_processor,
+                    seed=-1,
+                    **kwargs):
+                outputs.append(token.item())
+                yield tokenizer.decode(outputs, skip_special_tokens=True, errors='ignore')
+
+        return stream_generator()
+
+    def generate(
+        self,
+        inputs: Optional[torch.Tensor] = None,
+        generation_config: Optional[GenerationConfig] = None,
+        logits_processor: Optional[LogitsProcessorList] = None,
+        stopping_criteria: Optional[StoppingCriteriaList] = None,
+        prefix_allowed_tokens_fn: Optional[
+            Callable[[int, torch.Tensor], List[int]]
+        ] = None,
+        synced_gpus: Optional[bool] = None,
+        assistant_model: Optional["PreTrainedModel"] = None,
+        streamer: Optional["BaseStreamer"] = None,
+        **kwargs,
+    ) -> Union[GenerateOutput, torch.LongTensor]:
+        generation_config = generation_config if generation_config is not None else self.generation_config
+
+        # Process stop_words_ids.
+        stop_words_ids = kwargs.pop("stop_words_ids", None)
+        if stop_words_ids is None and generation_config is not None:
+            stop_words_ids = getattr(generation_config, "stop_words_ids", None)
+        if stop_words_ids is None:
+            stop_words_ids = getattr(generation_config, "stop_words_ids", None)
+
+        if stop_words_ids is not None:
+            stop_words_logits_processor = StopWordsLogitsProcessor(
+                stop_words_ids=stop_words_ids,
+                eos_token_id=generation_config.eos_token_id,
+            )
+            if logits_processor is None:
+                logits_processor = LogitsProcessorList([stop_words_logits_processor])
+            else:
+                logits_processor.append(stop_words_logits_processor)
+
+        return super().generate(
+            inputs,
+            generation_config=generation_config,
+            logits_processor=logits_processor,
+            stopping_criteria=stopping_criteria,
+            prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
+            synced_gpus=synced_gpus,
+            assistant_model=assistant_model,
+            streamer=streamer,
+            **kwargs,
+        )
+
+
+class RotaryEmbedding(torch.nn.Module):
+    def __init__(self, dim, base=10000):
+        super().__init__()
+        self.dim = dim
+        self.base = base
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        if importlib.util.find_spec("einops") is None:
+            raise RuntimeError("einops is required for Rotary Embedding")
+
+        self._rotary_pos_emb_cache = None
+        self._seq_len_cached = 0
+        self._ntk_alpha_cached = 1.0
+        self._ntk_alpha_cached_list = [1.0]
+
+    def update_rotary_pos_emb_cache(self, max_seq_len, offset=0, ntk_alpha=1.0):
+        seqlen = max_seq_len + offset
+        if seqlen > self._seq_len_cached or ntk_alpha != self._ntk_alpha_cached:
+            base = self.base * ntk_alpha ** (self.dim / (self.dim - 2))
+            self.inv_freq = 1.0 / (
+                base
+                ** (
+                    torch.arange(0, self.dim, 2, device=self.inv_freq.device).float()
+                    / self.dim
+                )
+            )
+            self._seq_len_cached = max(2 * seqlen, 16)
+            self._ntk_alpha_cached = ntk_alpha
+            seq = torch.arange(self._seq_len_cached, device=self.inv_freq.device)
+            freqs = torch.outer(seq.type_as(self.inv_freq), self.inv_freq)
+
+            emb = torch.cat((freqs, freqs), dim=-1)
+            from einops import rearrange
+
+            emb = rearrange(emb, "n d -> 1 n 1 d")
+
+            cos, sin = emb.cos(), emb.sin()
+            self._rotary_pos_emb_cache = [cos, sin]
+
+    def forward(self, max_seq_len, offset=0, ntk_alpha=1.0):
+        self.update_rotary_pos_emb_cache(max_seq_len, offset, ntk_alpha)
+        cos, sin = self._rotary_pos_emb_cache
+        return [cos[:, offset : offset + max_seq_len], sin[:, offset : offset + max_seq_len]]
+
+
+def _rotate_half(x):
+    from einops import rearrange
+
+    x = rearrange(x, "... (j d) -> ... j d", j=2)
+    x1, x2 = x.unbind(dim=-2)
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(t, freqs):
+    cos, sin = freqs
+    if apply_rotary_emb_func is not None and t.is_cuda:
+        t_ = t.float()
+        cos = cos.squeeze(0).squeeze(1)[:, : cos.shape[-1] // 2]
+        sin = sin.squeeze(0).squeeze(1)[:, : sin.shape[-1] // 2]
+        output = apply_rotary_emb_func(t_, cos, sin).type_as(t)
+        return output
+    else:
+        rot_dim = freqs[0].shape[-1]
+        cos, sin = freqs
+        t_, t_pass_ = t[..., :rot_dim], t[..., rot_dim:]
+        t_ = t_.float()
+        t_pass_ = t_pass_.float()
+        t_ = (t_ * cos) + (_rotate_half(t_) * sin)
+        return torch.cat((t_, t_pass_), dim=-1).type_as(t)
+
+
+class RMSNorm(torch.nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.eps = eps
+        self.weight = nn.Parameter(torch.ones(dim))
+
+    def _norm(self, x):
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+    def forward(self, x):
+        if rms_norm is not None and x.is_cuda:
+            return rms_norm(x, self.weight, self.eps)
+        else:
+            output = self._norm(x.float()).type_as(x)
+            return output * self.weight

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6415987961a707da763cd76268de56c086d072d0962bdb11fab05212c5398160
+size 7210852530

quantize_config.json

@@ -0,0 +1,11 @@
+{
+  "bits": 4,
+  "group_size": 128,
+  "damp_percent": 0.01,
+  "desc_act": false,
+  "static_groups": false,
+  "sym": true,
+  "true_sequential": true,
+  "model_name_or_path": null,
+  "model_file_base_name": null
+}

qwen_generation_utils.py

@@ -0,0 +1,431 @@
+# Copyright (c) Alibaba Cloud.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Generation support."""
+
+from typing import Tuple, List, Union, Iterable, Dict
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from transformers import PreTrainedTokenizer
+from transformers import logging
+from transformers.generation import LogitsProcessor
+
+logger = logging.get_logger(__name__)
+
+# Types.
+HistoryType = List[Tuple[str, str]]
+TokensType = List[int]
+BatchTokensType = List[List[int]]
+
+
+def pad_batch(batch: BatchTokensType, pad_id: int, seq_length: int) -> BatchTokensType:
+    for tokens in batch:
+        context_length = len(tokens)
+        if context_length < seq_length:
+            tokens.extend([pad_id] * (seq_length - context_length))
+    return batch
+
+
+def get_ltor_masks_and_position_ids(
+    data,
+    eod_token,
+    reset_position_ids,
+    reset_attention_mask,
+    eod_mask_loss,
+):
+    """Build masks and position id for left to right model."""
+
+    # Extract batch size and sequence length.
+    micro_batch_size, seq_length = data.size()
+
+    # Attention mask (lower triangular).
+    if reset_attention_mask:
+        att_mask_batch = micro_batch_size
+    else:
+        att_mask_batch = 1
+    attention_mask = torch.tril(
+        torch.ones((att_mask_batch, seq_length, seq_length), device=data.device)
+    ).view(att_mask_batch, 1, seq_length, seq_length)
+
+    # Loss mask.
+    loss_mask = torch.ones(data.size(), dtype=torch.float, device=data.device)
+    if eod_mask_loss:
+        loss_mask[data == eod_token] = 0.0
+
+    # Position ids.
+    position_ids = torch.arange(seq_length, dtype=torch.long, device=data.device)
+    position_ids = position_ids.unsqueeze(0).expand_as(data)
+    # We need to clone as the ids will be modifed based on batch index.
+    if reset_position_ids:
+        position_ids = position_ids.clone()
+
+    if reset_position_ids or reset_attention_mask:
+        # Loop through the batches:
+        for b in range(micro_batch_size):
+
+            # Find indecies where EOD token is.
+            eod_index = position_ids[b, data[b] == eod_token]
+            # Detach indecies from positions if going to modify positions.
+            if reset_position_ids:
+                eod_index = eod_index.clone()
+
+            # Loop through EOD indecies:
+            prev_index = 0
+            for j in range(eod_index.size()[0]):
+                i = eod_index[j]
+                # Mask attention loss.
+                if reset_attention_mask:
+                    attention_mask[b, 0, (i + 1) :, : (i + 1)] = 0
+                # Reset positions.
+                if reset_position_ids:
+                    position_ids[b, (i + 1) :] -= i + 1 - prev_index
+                    prev_index = i + 1
+
+    # Convert attention mask to binary:
+    attention_mask = attention_mask < 0.5
+
+    return attention_mask, loss_mask, position_ids
+
+
+def get_batch(context_tokens: torch.LongTensor, eod_id: int):
+    """Generate batch from context tokens."""
+    # Move to GPU.
+    tokens = context_tokens.contiguous().to(context_tokens.device)
+    # Get the attention mask and postition ids.
+    attention_mask, _, position_ids = get_ltor_masks_and_position_ids(
+        tokens,
+        eod_id,
+        reset_position_ids=False,
+        reset_attention_mask=False,
+        eod_mask_loss=False,
+    )
+    return tokens, attention_mask, position_ids
+
+
+def get_stop_words_ids(chat_format, tokenizer):
+    if chat_format == "raw":
+        stop_words_ids = [tokenizer.encode("Human:"), [tokenizer.eod_id]]
+    elif chat_format == "chatml":
+        stop_words_ids = [[tokenizer.im_end_id], [tokenizer.im_start_id]]
+    else:
+        raise NotImplementedError(f"Unknown chat format {chat_format!r}")
+    return stop_words_ids
+
+
+def make_context(
+    tokenizer: PreTrainedTokenizer,
+    query: str,
+    history: List[Tuple[str, str]] = None,
+    system: str = "",
+    max_window_size: int = 6144,
+    chat_format: str = "chatml",
+):
+    audio_info = None
+    if history is None:
+        history = []
+
+    if chat_format == "chatml":
+        im_start, im_end = "<|im_start|>", "<|im_end|>"
+        im_start_tokens = [tokenizer.im_start_id]
+        im_end_tokens = [tokenizer.im_end_id]
+        nl_tokens = tokenizer.encode("\n")
+
+        def _tokenize_str(role, content):
+            # import ipdb; ipdb.set_trace()
+            audio_info = tokenizer.process_audio(content)
+            return f"{role}\n{content}", tokenizer.encode(
+                role, allowed_special=set(tokenizer.AUDIO_ST), audio_info=audio_info
+            ) + nl_tokens + tokenizer.encode(content, allowed_special=set(tokenizer.AUDIO_ST), audio_info=audio_info),audio_info
+
+        system_text, system_tokens_part, audio_info = _tokenize_str("system", system)
+        system_tokens = im_start_tokens + system_tokens_part + im_end_tokens
+
+        raw_text = ""
+        context_tokens = []
+
+        for turn_query, turn_response in reversed(history):
+            query_text, query_tokens_part, _ = _tokenize_str("user", turn_query)
+            query_tokens = im_start_tokens + query_tokens_part + im_end_tokens
+            if turn_response is not None:
+                response_text, response_tokens_part, _ = _tokenize_str(
+                    "assistant", turn_response
+                )
+                response_tokens = im_start_tokens + response_tokens_part + im_end_tokens
+
+                next_context_tokens = nl_tokens + query_tokens + nl_tokens + response_tokens
+                prev_chat = (
+                    f"\n{im_start}{query_text}{im_end}\n{im_start}{response_text}{im_end}"
+                )
+            else:
+                next_context_tokens = nl_tokens + query_tokens + nl_tokens
+                prev_chat = f"\n{im_start}{query_text}{im_end}\n"
+
+            current_context_size = (
+                len(system_tokens) + len(next_context_tokens) + len(context_tokens)
+            )
+            if current_context_size < max_window_size:
+                context_tokens = next_context_tokens + context_tokens
+                raw_text = prev_chat + raw_text
+            else:
+                break
+
+        context_tokens = system_tokens + context_tokens
+        raw_text = f"{im_start}{system_text}{im_end}" + raw_text
+        context_tokens += (
+            nl_tokens
+            + im_start_tokens
+            + _tokenize_str("user", query)[1]
+            + im_end_tokens
+            + nl_tokens
+            + im_start_tokens
+            + tokenizer.encode("assistant")
+            + nl_tokens
+        )
+        raw_text += f"\n{im_start}user\n{query}{im_end}\n{im_start}assistant\n"
+        audio_info = tokenizer.process_audio(raw_text)
+
+    elif chat_format == "raw":
+        raw_text = query
+        context_tokens = tokenizer.encode(raw_text)
+    else:
+        raise NotImplementedError(f"Unknown chat format {chat_format!r}")
+
+    return raw_text, context_tokens, audio_info
+
+
+def _decode_default(
+    tokens: List[int],
+    *,
+    stop_words: List[str],
+    eod_words: List[str],
+    tokenizer: PreTrainedTokenizer,
+    raw_text_len: int,
+    verbose: bool = False,
+    return_end_reason: bool = False,
+    errors: str='replace',
+    audio_info:Dict = None
+):
+    kwargs = {"audio_info": audio_info}
+    trim_decode_tokens = tokenizer.decode(tokens, errors=errors, **kwargs)[raw_text_len:]
+    if verbose:
+        print("\nRaw Generate: ", trim_decode_tokens)
+
+    end_reason = f"Gen length {len(tokens)}"
+    for stop_word in stop_words:
+        trim_decode_tokens = trim_decode_tokens.replace(stop_word, "").strip()
+    for eod_word in eod_words:
+        if eod_word in trim_decode_tokens:
+            end_reason = f"Gen {eod_word!r}"
+        trim_decode_tokens = trim_decode_tokens.split(eod_word)[0]
+    trim_decode_tokens = trim_decode_tokens.strip()
+    if verbose:
+        print("\nEnd Reason:", end_reason)
+        print("\nGenerate: ", trim_decode_tokens)
+
+    if return_end_reason:
+        return trim_decode_tokens, end_reason
+    else:
+        return trim_decode_tokens
+
+
+def _decode_chatml(
+    tokens: List[int],
+    *,
+    stop_words: List[str],
+    eod_token_ids: List[int],
+    tokenizer: PreTrainedTokenizer,
+    raw_text_len: int,
+    context_length: int,
+    verbose: bool = False,
+    return_end_reason: bool = False,
+    errors: str='replace',
+    audio_info: Dict = None
+):
+    kwargs = {"audio_info": audio_info}
+    end_reason = f"Gen length {len(tokens)}"
+    eod_token_idx = context_length
+    for eod_token_idx in range(context_length, len(tokens)):
+        if tokens[eod_token_idx] in eod_token_ids:
+            end_reason = f"Gen {tokenizer.decode([tokens[eod_token_idx]],**kwargs)!r}"
+            break
+
+    trim_decode_tokens = tokenizer.decode(tokens[:eod_token_idx], errors=errors, **kwargs)[raw_text_len:]
+    if verbose:
+        print("\nRaw Generate w/o EOD:", tokenizer.decode(tokens, errors=errors, **kwargs)[raw_text_len:])
+        print("\nRaw Generate:", trim_decode_tokens)
+        print("\nEnd Reason:", end_reason)
+    for stop_word in stop_words:
+        trim_decode_tokens = trim_decode_tokens.replace(stop_word, "").strip()
+    trim_decode_tokens = trim_decode_tokens.strip()
+    if verbose:
+        print("\nGenerate:", trim_decode_tokens)
+
+    if return_end_reason:
+        return trim_decode_tokens, end_reason
+    else:
+        return trim_decode_tokens
+
+
+def decode_tokens(
+    tokens: Union[torch.LongTensor, TokensType],
+    tokenizer: PreTrainedTokenizer,
+    raw_text_len: int,
+    context_length: int,
+    chat_format: str,
+    verbose: bool = False,
+    return_end_reason: bool = False,
+    errors: str="replace",
+    audio_info: Dict = None
+) -> str:
+    if torch.is_tensor(tokens):
+        tokens = tokens.cpu().numpy().tolist()
+
+    if chat_format == "chatml":
+        return _decode_chatml(
+            tokens,
+            stop_words=[],
+            eod_token_ids=[tokenizer.im_start_id, tokenizer.im_end_id],
+            tokenizer=tokenizer,
+            raw_text_len=raw_text_len,
+            context_length=context_length,
+            verbose=verbose,
+            return_end_reason=return_end_reason,
+            errors=errors,
+            audio_info=audio_info
+        )
+    elif chat_format == "raw":
+        return _decode_default(
+            tokens,
+            stop_words=["<|endoftext|>"],
+            eod_words=["<|endoftext|>"],
+            tokenizer=tokenizer,
+            raw_text_len=raw_text_len,
+            verbose=verbose,
+            return_end_reason=return_end_reason,
+            errors=errors,
+            audio_info=audio_info
+        )
+    else:
+        raise NotImplementedError(f"Unknown chat format {chat_format!r}")
+
+
+class StopWordsLogitsProcessor(LogitsProcessor):
+    """
+    :class:`transformers.LogitsProcessor` that enforces that when specified sequences appear, stop geration.
+
+    Args:
+        stop_words_ids (:obj:`List[List[int]]`):
+            List of list of token ids of stop ids. In order to get the tokens of the words
+            that should not appear in the generated text, use :obj:`tokenizer(bad_word,
+            add_prefix_space=True).input_ids`.
+        eos_token_id (:obj:`int`):
+            The id of the `end-of-sequence` token.
+    """
+
+    def __init__(self, stop_words_ids: Iterable[Iterable[int]], eos_token_id: int):
+
+        if not isinstance(stop_words_ids, List) or len(stop_words_ids) == 0:
+            raise ValueError(
+                f"`stop_words_ids` has to be a non-emtpy list, but is {stop_words_ids}."
+            )
+        if any(not isinstance(bad_word_ids, list) for bad_word_ids in stop_words_ids):
+            raise ValueError(
+                f"`stop_words_ids` has to be a list of lists, but is {stop_words_ids}."
+            )
+        if any(
+            any(
+                (not isinstance(token_id, (int, np.integer)) or token_id < 0)
+                for token_id in stop_word_ids
+            )
+            for stop_word_ids in stop_words_ids
+        ):
+            raise ValueError(
+                f"Each list in `stop_words_ids` has to be a list of positive integers, but is {stop_words_ids}."
+            )
+
+        self.stop_words_ids = list(
+            filter(
+                lambda bad_token_seq: bad_token_seq != [eos_token_id], stop_words_ids
+            )
+        )
+        self.eos_token_id = eos_token_id
+        for stop_token_seq in self.stop_words_ids:
+            assert (
+                len(stop_token_seq) > 0
+            ), "Stop words token sequences {} cannot have an empty list".format(
+                stop_words_ids
+            )
+
+    def __call__(
+        self, input_ids: torch.LongTensor, scores: torch.FloatTensor
+    ) -> torch.FloatTensor:
+        stopped_samples = self._calc_stopped_samples(input_ids)
+        for i, should_stop in enumerate(stopped_samples):
+            if should_stop:
+                scores[i, self.eos_token_id] = float(2**15)
+        return scores
+
+    def _tokens_match(self, prev_tokens: torch.LongTensor, tokens: List[int]) -> bool:
+        if len(tokens) == 0:
+            # if bad word tokens is just one token always ban it
+            return True
+        elif len(tokens) > len(prev_tokens):
+            # if bad word tokens are longer then prev input_ids they can't be equal
+            return False
+        elif prev_tokens[-len(tokens) :].tolist() == tokens:
+            # if tokens match
+            return True
+        else:
+            return False
+
+    def _calc_stopped_samples(self, prev_input_ids: Iterable[int]) -> Iterable[int]:
+        stopped_samples = []
+        for prev_input_ids_slice in prev_input_ids:
+            match = False
+            for stop_token_seq in self.stop_words_ids:
+                if self._tokens_match(prev_input_ids_slice, stop_token_seq):
+                    # if tokens do not match continue
+                    match = True
+                    break
+            stopped_samples.append(match)
+
+        return stopped_samples
+
+
+def top_k_logits(logits, top_k=0, top_p=0.0, filter_value=-float("Inf")):
+    """This function has been mostly taken from huggingface conversational
+    ai code at
+        https://medium.com/huggingface/how-to-build-a-state-of-the-art-
+             conversational-ai-with-transfer-learning-2d818ac26313"""
+
+    if top_k > 0:
+        # Remove all tokens with a probability less than the
+        # last token of the top-k
+        indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
+        logits[indices_to_remove] = filter_value
+
+    if top_p > 0.0:
+        # Cconvert to 1D
+        sorted_logits, sorted_indices = torch.sort(logits, descending=True, dim=-1)
+        cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
+
+        # Remove tokens with cumulative probability above the threshold
+        sorted_indices_to_remove = cumulative_probs > top_p
+        # Shift the indices to the right to keep also the first token
+        # above the threshold
+        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+        sorted_indices_to_remove[..., 0] = 0
+        for i in range(sorted_indices.size(0)):
+            indices_to_remove = sorted_indices[i][sorted_indices_to_remove[i]]
+            logits[i][indices_to_remove] = filter_value
+
+    return logits
+
+
+def switch(val1, val2, boolean):
+    boolean = boolean.type_as(val1)
+    return (1 - boolean) * val1 + boolean * val2

special_tokens_map.json

@@ -0,0 +1 @@
+{}

tokenization_qwen.py

@@ -0,0 +1,579 @@
+# Copyright (c) Alibaba Cloud.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Tokenization classes for QWen."""
+
+import base64
+import logging
+import os
+import re
+import itertools
+
+import requests
+import unicodedata
+from typing import Collection, Dict, List, Set, Tuple, Union, Any, Callable, Optional
+
+import tiktoken
+import numpy as np
+
+from transformers import PreTrainedTokenizer, AddedToken
+from transformers.utils import try_to_load_from_cache
+from transformers.tokenization_utils_base import BatchEncoding, PaddingStrategy, TruncationStrategy, \
+    TextInput, TextInputPair, PreTokenizedInput, PreTokenizedInputPair, TensorType, EncodedInput, EncodedInputPair
+
+import matplotlib.colors as mcolors
+from matplotlib.font_manager import FontProperties
+from .audio import *
+
+logger = logging.getLogger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "qwen.tiktoken", "ttf": "SimSun.ttf"}
+
+PAT_STR = r"""(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+"""
+ENDOFTEXT = "<|endoftext|>"
+IMSTART = "<|im_start|>"
+IMEND = "<|im_end|>"
+# as the default behavior is changed to allow special tokens in
+# regular texts, the surface forms of special tokens need to be
+# as different as possible to minimize the impact
+EXTRAS = tuple((f"<|extra_{i}|>" for i in range(205)))
+SPECIAL_TOKENS = (
+                     ENDOFTEXT,
+                     IMSTART,
+                     IMEND,
+                 ) + EXTRAS
+
+LANGUAGES = {
+    "en": "english",
+    "zh": "chinese",
+    "de": "german",
+    "es": "spanish",
+    "ko": "korean",
+    "fr": "french",
+    "ja": "japanese",
+    "it": "italian",
+}
+
+
+def _load_tiktoken_bpe(tiktoken_bpe_file: str) -> Dict[bytes, int]:
+    with open(tiktoken_bpe_file, "rb") as f:
+        contents = f.read()
+    return {
+        base64.b64decode(token): int(rank)
+        for token, rank in (line.split() for line in contents.splitlines() if line)
+    }
+
+
+def _list_find(
+        input_list: List[Any],
+        candidates: Tuple[Any],
+        start: int = 0,
+):
+    for i in range(start, len(input_list)):
+        if input_list[i] in candidates:
+            return i
+    return -1
+
+
+def _replace_closed_tag(
+        input_tokens: List[Any],
+        start_tags: Union[Any, Tuple[Any]],
+        end_tags: Union[Any, Tuple[Any]],
+        inclusive_replace_func: Callable,
+        exclusive_replace_func: Callable = lambda x: x,
+        audio_info: Dict = None
+):
+    if isinstance(start_tags, (str, int)):
+        start_tags = (start_tags,)
+    if isinstance(end_tags, (str, int)):
+        end_tags = (end_tags,)
+    assert len(start_tags) == len(end_tags)
+
+    output_tokens = []
+    end = 0
+    audio_idx = 0
+    while True:
+        start = _list_find(input_tokens, start_tags, end)
+        if start == -1:
+            break
+        output_tokens.extend(exclusive_replace_func(input_tokens[end: start]))
+        tag_idx = start_tags.index(input_tokens[start])
+        end = _list_find(input_tokens, (end_tags[tag_idx],), start)
+        if end == -1:
+            raise ValueError("Unclosed audio token")
+        output_tokens.extend(inclusive_replace_func(input_tokens[start: end + 1], audio_info, audio_idx))
+        end += 1
+        audio_idx += 1
+    output_tokens.extend(exclusive_replace_func(input_tokens[end:]))
+    return output_tokens
+
+
+class QWenTokenizer(PreTrainedTokenizer):
+    """QWen tokenizer."""
+
+    vocab_files_names = VOCAB_FILES_NAMES
+
+    def __init__(
+            self,
+            vocab_file,
+            errors="replace",
+            audio_start_tag='<audio>',
+            audio_end_tag='</audio>',
+            **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.audio_start_tag = audio_start_tag
+        self.audio_end_tag = audio_end_tag
+        self.audio_pad_tag = "[[[AUDIO:modality]]]"
+
+        self.AUDIO_ST = (
+            '[[[AUDIO:modality]]]',
+            # Transcription Tag
+            "<|startoftranscript|>",  # Transcription
+            "<|startofanalysis|>",  # Analysis
+            # Task Tag
+            "<|translate|>",
+            "<|transcribe|>",
+            "<|caption|>",
+            "<|keyword|>",
+            # Language Tag
+            "<|unknown|>",  # unknown language
+            *[f"<|{lang}|>" for lang in LANGUAGES.keys()],
+            "<|zh_tr|>",  # tranditional Chinese
+            # Timestamps Tag
+            "<|notimestamps|>",
+            "<|sil|>",
+            "<|timestamps|>",
+            *[f"<|{i * 0.01:.2f}|>" for i in range(3001)],  # timestamps 0.00-30.00
+            # Output Instruction
+            "<|caption_audiocaps|>",  # Audiocaps caption style
+            "<|caption_clotho|>",  # Clotho caption style
+            "<|audioset_ontology|>",  # Audioset ontology style
+            "<|caption_plain|>",  # plain caption
+            "<|itn|>",  # inversed text normalized
+            "<|wo_itn|>",  # without inversed text normalized
+            "<|startofentityvalue|>",
+            "<|endofentityvalue|>",
+            "<|startofentitytype|>",
+            "<|endofentitytype|>",
+            "<|named_entity_recognition|>",  # named entity recognition task
+            "<|audio_grounding|>",
+            "<|startofword|>",
+            "<|endofword|>",
+            "<|delim|>",  # delimiter of timestamps pair in audio grounding
+            "<|emotion_recognition|>",  # emotion recognition
+            "<|music_description|>",  # music description
+            "<|note_analysis|>",  # note analysis
+            "<|pitch|>",  # note analysis: pitch
+            *[f"<|midi_pitch_{i}|>" for i in range(128)],  # midi pitch 0-127
+            "<|velocity|>",  # note analysis: velocity
+            *[f"<|midi_velocity_{i}|>" for i in range(128)],  # midi velocity 0-127
+            "<|sonic|>",  # note analysis:  sonic
+            "<|instrument|>",  # note analysis:  instrument
+            "<|speaker_meta|>",  # meta information of speaker
+            "<|song_meta|>",  # meta information of song
+            "<|question|>",  # AQA: question
+            "<|answer|>",  # AQA: answer
+            "<|choice|>",  # AQA: answer choice
+            "<|scene|>",  # scene recognition
+            "<|event|>",  # sound event
+            "<|vocal_classification|>",  # vocal classification
+            "<|speech_understanding|>",  # speech language understanding
+            "<|scenario|>",  # speech language understanding: scenario
+            "<|action|>",  # speech language understanding: action
+            "<|entities|>",  # speech language understanding: entities
+            "<|speech_edit|>",  # speech edit
+            audio_start_tag,
+            audio_end_tag
+        )
+
+        self.errors = errors  # how to handle errors in decoding
+
+        self.mergeable_ranks = _load_tiktoken_bpe(vocab_file)  # type: dict[bytes, int]
+        self.special_tokens = {
+            token: index
+            for index, token in enumerate(
+                SPECIAL_TOKENS + self.AUDIO_ST, start=len(self.mergeable_ranks)
+
+            )
+        }
+        self.audio_start_id = self.special_tokens[self.audio_start_tag]
+        self.audio_end_id = self.special_tokens[self.audio_end_tag]
+        self.audio_pad_id = self.special_tokens[self.audio_pad_tag]
+        print(f"audio_start_id: {self.audio_start_id}, "
+              f"audio_end_id: {self.audio_end_id}, "
+              f"audio_pad_id: {self.audio_pad_id}.")
+
+        enc = tiktoken.Encoding(
+            "Qwen",
+            pat_str=PAT_STR,
+            mergeable_ranks=self.mergeable_ranks,
+            special_tokens=self.special_tokens,
+        )
+        assert (
+                len(self.mergeable_ranks) + len(self.special_tokens) == enc.n_vocab
+        ), f"{len(self.mergeable_ranks) + len(self.special_tokens)} != {enc.n_vocab} in encoding"
+
+        self.decoder = {
+            v: k for k, v in self.mergeable_ranks.items()
+        }  # type: dict[int, bytes|str]
+        self.decoder.update({v: k for k, v in self.special_tokens.items()})
+
+        self.tokenizer = enc  # type: tiktoken.Encoding
+
+        self.eod_id = self.tokenizer.eot_token
+        self.im_start_id = self.special_tokens[IMSTART]
+        self.im_end_id = self.special_tokens[IMEND]
+
+    def __getstate__(self):
+        # for pickle lovers
+        state = self.__dict__.copy()
+        del state['tokenizer']
+        return state
+
+    def __setstate__(self, state):
+        # tokenizer is not python native; don't pass it; rebuild it
+        self.__dict__.update(state)
+        enc = tiktoken.Encoding(
+            "Qwen",
+            pat_str=PAT_STR,
+            mergeable_ranks=self.mergeable_ranks,
+            special_tokens=self.special_tokens,
+        )
+        self.tokenizer = enc
+
+    def __len__(self) -> int:
+        return self.tokenizer.n_vocab
+
+    def get_vocab(self) -> Dict[bytes, int]:
+        return self.mergeable_ranks
+
+    def convert_tokens_to_ids(
+            self, tokens: Union[bytes, str, List[Union[bytes, str]]]
+    ) -> List[int]:
+        ids = []
+        if isinstance(tokens, (str, bytes)):
+            if tokens in self.special_tokens:
+                return self.special_tokens[tokens]
+            else:
+                return self.mergeable_ranks.get(tokens)
+        for token in tokens:
+            if token in self.special_tokens:
+                ids.append(self.special_tokens[token])
+            else:
+                ids.append(self.mergeable_ranks.get(token))
+        return ids
+
+    def _add_tokens(self, new_tokens: Union[List[str], List[AddedToken]], special_tokens: bool = False) -> int:
+        if not special_tokens and new_tokens:
+            raise ValueError('Adding regular tokens is not supported')
+        for token in new_tokens:
+            surface_form = token.content if isinstance(token, AddedToken) else token
+            if surface_form not in SPECIAL_TOKENS  + self.AUDIO_ST:
+                raise ValueError('Adding unknown special tokens is not supported')
+        return 0
+
+    def save_vocabulary(self, save_directory: str, **kwargs) -> Tuple[str]:
+        """
+        Save only the vocabulary of the tokenizer (vocabulary).
+
+        Returns:
+            `Tuple(str)`: Paths to the files saved.
+        """
+        file_path = os.path.join(save_directory, "qwen.tiktoken")
+        with open(file_path, "w", encoding="utf8") as w:
+            for k, v in self.mergeable_ranks.items():
+                line = base64.b64encode(k).decode("utf8") + " " + str(v) + "\n"
+                w.write(line)
+        return (file_path,)
+
+    def tokenize(
+            self,
+            text: str,
+            allowed_special: Union[Set, str] = "all",
+            disallowed_special: Union[Collection, str] = (),
+            audio_info: Dict = None,
+            **kwargs,
+    ) -> List[Union[bytes, str]]:
+        """
+        Converts a string in a sequence of tokens.
+
+        Args:
+            text (`str`):
+                The sequence to be encoded.
+            allowed_special (`Literal["all"]` or `set`):
+                The surface forms of the tokens to be encoded as special tokens in regular texts.
+                Default to "all".
+            disallowed_special (`Literal["all"]` or `Collection`):
+                The surface forms of the tokens that should not be in regular texts and trigger errors.
+                Default to an empty tuple.
+
+            kwargs (additional keyword arguments, *optional*):
+                Will be passed to the underlying model specific encode method.
+
+        Returns:
+            `List[bytes|str]`: The list of tokens.
+        """
+        tokens = []
+        text = unicodedata.normalize("NFC", text)
+
+        # this implementation takes a detour: text -> token id -> token surface forms
+        for t in self.tokenizer.encode(
+                text, allowed_special=allowed_special, disallowed_special=disallowed_special
+        ):
+            tokens.append(self.decoder[t])
+
+        def _encode_audiourl(audio_tokens, audio_info, audio_idx):
+            assert audio_tokens[0] == self.audio_start_tag and audio_tokens[-1] == self.audio_end_tag
+            audio_token_span = audio_info['audio_span_tokens'][audio_idx]
+            out_audio_tokens = [self.audio_start_tag] + [self.audio_pad_tag] * (audio_token_span - 2) + [
+                self.audio_end_tag]
+            return out_audio_tokens
+
+        return _replace_closed_tag(tokens, self.audio_start_tag, self.audio_end_tag, _encode_audiourl,
+                                   audio_info=audio_info)
+
+    def _batch_encode_plus(
+            self,
+            batch_text_or_text_pairs: Union[
+                List[TextInput],
+                List[TextInputPair],
+                List[PreTokenizedInput],
+                List[PreTokenizedInputPair],
+                List[EncodedInput],
+                List[EncodedInputPair],
+            ],
+            add_special_tokens: bool = True,
+            padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD,
+            truncation_strategy: TruncationStrategy = TruncationStrategy.DO_NOT_TRUNCATE,
+            max_length: Optional[int] = None,
+            stride: int = 0,
+            is_split_into_words: bool = False,
+            pad_to_multiple_of: Optional[int] = None,
+            return_tensors: Optional[Union[str, TensorType]] = None,
+            return_token_type_ids: Optional[bool] = None,
+            return_attention_mask: Optional[bool] = None,
+            return_overflowing_tokens: bool = False,
+            return_special_tokens_mask: bool = False,
+            return_offsets_mapping: bool = False,
+            return_length: bool = False,
+            verbose: bool = True,
+            **kwargs,
+    ) -> BatchEncoding:
+
+        def get_input_ids(text):
+            if isinstance(text, str):
+                tokens = self.tokenize(text, **kwargs)
+                return self.convert_tokens_to_ids(tokens)
+            elif isinstance(text, (list, tuple)) and len(text) > 0 and isinstance(text[0], str):
+                if is_split_into_words:
+                    tokens = list(
+                        itertools.chain(*(self.tokenize(t, is_split_into_words=True, **kwargs) for t in text))
+                    )
+                    return self.convert_tokens_to_ids(tokens)
+                else:
+                    return self.convert_tokens_to_ids(text)
+            elif isinstance(text, (list, tuple)) and len(text) > 0 and isinstance(text[0], int):
+                return text
+            else:
+                raise ValueError(
+                    "Input is not valid. Should be a string, a list/tuple of strings or a list/tuple of integers."
+                )
+
+        if return_offsets_mapping:
+            raise NotImplementedError(
+                "return_offset_mapping is not available when using Python tokenizers. "
+                "To use this feature, change your tokenizer to one deriving from "
+                "transformers.PreTrainedTokenizerFast."
+            )
+
+        input_ids = []
+        audio_info = kwargs.pop("audio_info", None)
+        for pair_id in range(len(batch_text_or_text_pairs)):
+            kwargs['audio_info'] = audio_info[pair_id]
+            ids_or_pair_ids = batch_text_or_text_pairs[pair_id]
+            # for ids_or_pair_ids in batch_text_or_text_pairs:
+            if not isinstance(ids_or_pair_ids, (list, tuple)):
+                ids, pair_ids = ids_or_pair_ids, None
+            elif is_split_into_words and not isinstance(ids_or_pair_ids[0], (list, tuple)):
+                ids, pair_ids = ids_or_pair_ids, None
+            else:
+                ids, pair_ids = ids_or_pair_ids
+
+            first_ids = get_input_ids(ids)
+            second_ids = get_input_ids(pair_ids) if pair_ids is not None else None
+            input_ids.append((first_ids, second_ids))
+
+        batch_outputs = self._batch_prepare_for_model(
+            input_ids,
+            add_special_tokens=add_special_tokens,
+            padding_strategy=padding_strategy,
+            truncation_strategy=truncation_strategy,
+            max_length=max_length,
+            stride=stride,
+            pad_to_multiple_of=pad_to_multiple_of,
+            return_attention_mask=return_attention_mask,
+            return_token_type_ids=return_token_type_ids,
+            return_overflowing_tokens=return_overflowing_tokens,
+            return_special_tokens_mask=return_special_tokens_mask,
+            return_length=return_length,
+            return_tensors=return_tensors,
+            verbose=verbose,
+        )
+
+        return BatchEncoding(batch_outputs)
+
+    def convert_tokens_to_string(self, tokens: List[Union[bytes, str]]) -> str:
+        """
+        Converts a sequence of tokens in a single string.
+        """
+        text = ""
+        temp = b""
+        for t in tokens:
+            if isinstance(t, str):
+                if temp:
+                    text += temp.decode("utf-8", errors=self.errors)
+                    temp = b""
+                text += t
+            elif isinstance(t, bytes):
+                temp += t
+            else:
+                raise TypeError("token should only be of type types or str")
+        if temp:
+            text += temp.decode("utf-8", errors=self.errors)
+        return text
+
+    @property
+    def vocab_size(self):
+        return self.tokenizer.n_vocab
+
+    def _convert_id_to_token(self, index: int) -> Union[bytes, str]:
+        """Converts an id to a token, special tokens included"""
+        if index in self.decoder:
+            return self.decoder[index]
+        raise ValueError("unknown ids")
+
+    def _convert_token_to_id(self, token: Union[bytes, str]) -> int:
+        """Converts a token to an id using the vocab, special tokens included"""
+        if token in self.special_tokens:
+            return self.special_tokens[token]
+        if token in self.mergeable_ranks:
+            return self.mergeable_ranks[token]
+        raise ValueError("unknown token")
+
+    def _tokenize(self, text: str, **kwargs):
+        """
+        Converts a string in a sequence of tokens (string), using the tokenizer. Split in words for word-based
+        vocabulary or sub-words for sub-word-based vocabularies (BPE/SentencePieces/WordPieces).
+
+        Do NOT take care of added tokens.
+        """
+        raise NotImplementedError
+
+    def _decode(
+            self,
+            token_ids: Union[int, List[int]],
+            skip_special_tokens: bool = False,
+            errors: str = None,
+            **kwargs,
+    ) -> str:
+        if isinstance(token_ids, int):
+            token_ids = [token_ids]
+        audio_info = kwargs.pop("audio_info", None)
+
+        def _decode_audiourl(audio_token_ids, audio_info, audio_idx):
+            assert audio_token_ids[0] == self.audio_start_id and audio_token_ids[-1] == self.audio_end_id
+            audio_url = audio_info["audio_urls"][audio_idx]
+            return [self.audio_start_id] + self.tokenizer.encode(audio_url) + [self.audio_end_id]
+
+        token_ids = _replace_closed_tag(token_ids, self.audio_start_id, self.audio_end_id, _decode_audiourl,
+                                        audio_info=audio_info)
+
+        if skip_special_tokens:
+            token_ids = [i for i in token_ids if i < self.eod_id]
+        return self.tokenizer.decode(token_ids, errors=errors or self.errors)
+
+    def to_list_format(self, text: str):
+        text = unicodedata.normalize("NFC", text)
+        token_ids = self.tokenizer.encode(
+            text, allowed_special=set(self.AUDIO_ST + (ENDOFTEXT,)))
+
+        def _encode_audio_info(tokens):
+            if len(tokens) == 0:
+                return []
+            if tokens[0] == self.audio_start_id and tokens[-1] == self.audio_end_id:
+                key = 'audio'
+            else:
+                _tobytes = lambda x: x.encode('utf-8') if isinstance(x, str) else x
+                return [{'text': b''.join(map(_tobytes, map(self.decoder.get, tokens))).decode('utf-8')}]
+            _tobytes = lambda x: x.encode('utf-8') if isinstance(x, str) else x
+            val = b''.join(map(_tobytes, map(self.decoder.get, tokens[1:-1]))).decode('utf-8')
+            return [{key: val}]
+
+        return _replace_closed_tag(
+            token_ids,
+            (self.audio_start_id),
+            (self.audio_end_id),
+            _encode_audio_info,
+            _encode_audio_info,
+        )
+
+    def from_list_format(self, list_format: List[Dict]):
+        text = ''
+        num_audios = 0
+        for ele in list_format:
+            if 'audio' in ele:
+                num_audios += 1
+                text += f'Audio {num_audios}:'
+                text += self.audio_start_tag + ele['audio'] + self.audio_end_tag
+                text += '\n'
+            elif 'text' in ele:
+                text += ele['text']
+            elif 'box' in ele:
+                if 'ref' in ele:
+                    text += self.ref_start_tag + ele['ref'] + self.ref_end_tag
+                for box in ele['box']:
+                    text += self.box_start_tag + '(%d,%d),(%d,%d)' % (box[0], box[1], box[2], box[3]) + self.box_end_tag
+            else:
+                raise ValueError("Unsupport element: " + str(ele))
+        return text
+
+    def extract_audio_urls(self, text):
+        pattern = rf"{self.audio_start_tag}(.*?){self.audio_end_tag}"
+        return re.findall(pattern, text)
+
+    def process_audio(self, text):
+        audio_urls = self.extract_audio_urls(text)
+        if len(audio_urls) > 0:
+            audios, audio_lens, audio_span_tokens = [], [], []
+            for audio_path in audio_urls:
+                if audio_path.startswith("http://") or audio_path.startswith("https://"):  # http
+                    data = bytes(requests.get(audio_path, stream=True).content)
+                    audio = load_bytesio_audio(data)
+                else:
+                    audio = load_audio(audio_path)
+                L = (audio.shape[0] if audio.shape[0] <= 480000 else 480000)  # max_length < 30s
+                mel_len = L // 160
+                audio = pad_or_trim(audio.flatten())
+                mel = log_mel_spectrogram(audio)
+                audio_len_after_cnn = get_T_after_cnn(mel_len)
+                audio_token_num = (audio_len_after_cnn - 2) // 2 + 1
+                audio_len = [audio_len_after_cnn, audio_token_num]
+                audios.append(mel)
+                audio_lens.append(audio_len)
+                audio_span_tokens.append(audio_token_num + 2)  # add audio bos eos
+            input_audio_lengths = torch.IntTensor(audio_lens)
+            input_audios = torch.stack(audios, dim=0)
+            return {"input_audios": input_audios,
+                    "input_audio_lengths": input_audio_lengths,
+                    "audio_span_tokens": audio_span_tokens,
+                    "audio_urls": audio_urls}
+        else:
+            return None
+
+
+
+
+

tokenizer_config.json

@@ -0,0 +1,11 @@
+{
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenization_qwen.QWenTokenizer",
+      null
+    ]
+  },
+  "clean_up_tokenization_spaces": true,
+  "model_max_length": 2048,
+  "tokenizer_class": "QWenTokenizer"
+}