diff --git a/.gitignore b/.gitignore index df3953d..cf99040 100644 --- a/.gitignore +++ b/.gitignore @@ -150,6 +150,9 @@ data/test .vscode *-my.* *.csv -test.py +test.* *users.json WeClone-audio/src/output*.wav +WeClone-audio/uv.lock +Spark-TTS-0.5B/ +uv.lock diff --git a/README.md b/README.md index 008ba47..9692bb7 100644 --- a/README.md +++ b/README.md @@ -2,7 +2,7 @@ ## 核心功能✨ - 💬 使用微信聊天记录微调LLM -- 🎙️ 使用微信语音消息结合大模型实现高质量声音克隆 👉[WeClone-audio](https://github.com/xming521/WeClone/tree/master/WeClone-audio) +- 🎙️ 使用微信语音消息➕0.5B大模型实现高质量声音克隆 👉[WeClone-audio](https://github.com/xming521/WeClone/tree/master/WeClone-audio) - 🔗 绑定到微信机器人，实现自己的数字分身 ## 特性与说明📋 @@ -21,44 +21,29 @@ 目前项目默认使用chatglm3-6b模型，LoRA方法对sft阶段微调，大约需要16GB显存。也可以使用[LLaMA Factory](https://github.com/hiyouga/LLaMA-Factory/blob/main/README_zh.md#%E6%A8%A1%E5%9E%8B)支持的其他模型和方法，占用显存更少，需要自行修改模板的system提示词等相关配置。 需要显存的估算值： -| 训练方法 | 精度 | 7B | 13B | 30B | 65B | 8x7B | -| ------- | ---- | ----- | ----- | ----- | ------ | ------ | -| 全参数 | 16 | 160GB | 320GB | 600GB | 1200GB | 900GB | -| 部分参数 | 16 | 20GB | 40GB | 120GB | 240GB | 200GB | -| LoRA | 16 | **16GB** | 32GB | 80GB | 160GB | 120GB | -| QLoRA | 8 | 10GB | 16GB | 40GB | 80GB | 80GB | -| QLoRA | 4 | 6GB | 12GB | 24GB | 48GB | 32GB | +| 方法 | 精度 | 7B | 14B | 30B | 70B | `x`B | +| ------------------------------- | ---- | ----- | ----- | ----- | ------ | ------- | +| Full (`bf16` or `fp16`) | 32 | 120GB | 240GB | 600GB | 1200GB | `18x`GB | +| Full (`pure_bf16`) | 16 | 60GB | 120GB | 300GB | 600GB | `8x`GB | +| Freeze/LoRA/GaLore/APOLLO/BAdam | 16 | 16GB | 32GB | 64GB | 160GB | `2x`GB | +| QLoRA | 8 | 10GB | 20GB | 40GB | 80GB | `x`GB | +| QLoRA | 4 | 6GB | 12GB | 24GB | 48GB | `x/2`GB | +| QLoRA | 2 | 4GB | 8GB | 16GB | 24GB | `x/4`GB | -### 软件要求 - -| 必需项 | 至少 | 推荐 | -| ------------ | ------- | --------- | -| python | 3.8 | 3.10 | -| torch | 1.13.1 | 2.2.1 | -| transformers | 4.37.2 | 4.38.1 | -| datasets | 2.14.3 | 2.17.1 | -| accelerate | 0.27.2 | 0.27.2 | -| peft | 0.9.0 | 0.9.0 | -| trl | 0.7.11 | 0.7.11 | - -| 可选项 | 至少 | 推荐 | -| ------------ | ------- | --------- | -| CUDA | 11.6 | 12.2 | -| deepspeed | 0.10.0 | 0.13.4 | -| bitsandbytes | 0.39.0 | 0.41.3 | -| flash-attn | 2.3.0 | 2.5.5 | ### 环境搭建 - +建议使用 [uv](https://docs.astral.sh/uv/)，这是一个非常快速的 Python 环境管理器。安装uv后，您可以使用以下命令创建一个新的Python环境并安装依赖项，注意这不包含xcodec（音频克隆）功能的依赖： ```bash git clone https://github.com/xming521/WeClone.git -conda create -n weclone python=3.10 -conda activate weclone cd WeClone -pip install -r requirements.txt +uv venv .venv --python=3.9 +source .venv/bin/activate +uv pip install --group main -e . ``` -训练以及推理相关配置统一在文件[settings.json](settings.json) +> [!NOTE] +> 训练以及推理相关配置统一在文件[settings.json](settings.json) + ### 数据准备 @@ -86,6 +71,7 @@ export USE_MODELSCOPE_HUB=1 # Windows 使用 `set USE_MODELSCOPE_HUB=1` git lfs install git clone https://www.modelscope.cn/ZhipuAI/chatglm3-6b.git ``` +魔搭社区的`modeling_chatglm.py`文件需要更换为Hugging Face的 ### 配置参数并微调模型 diff --git a/WeClone-audio/README.md b/WeClone-audio/README.md index b8eb98a..c9da00c 100644 --- a/WeClone-audio/README.md +++ b/WeClone-audio/README.md @@ -1,9 +1,16 @@ # WeClone-audio 模块 -WeClone-audio 是一个使用微信语音消息克隆声音的模块，使用 Llasa 模型实现高质量语音合成。 +WeClone-audio 是一个使用微信语音消息克隆声音的模块，使用模型实现高质量语音合成。 ### 显存需求 +**Spark-TTS** 推荐 +- **0.5B 模型**: 约 4GB 显存 + +**Llasa** - **3B 模型**: 约 16GB 显存 -- **1B 模型**: 约 9GB 显存 +- **1B 模型**: 约 9GB 显存 + + + ## 1. 导出微信语音数据 @@ -14,16 +21,14 @@ WeClone-audio 是一个使用微信语音消息克隆声音的模块，使用 Ll ### 1.2 环境配置 语音导出仅支持Windows环境 - -WeClone Audio使用uv作为包管理器，暂时独立于WeClone项目。请确保已安装uv。 +WeClone Audio使用uv作为包管理器。 ```bash # 为 PyWxDump 创建 Python 环境和安装依赖 # -cd ./WeClone-audio uv venv .venv-wx --python=3.9 source .venv-wx/bin/activate # 安装 wx 依赖组 -uv pip install -e '.[wx]' +uv pip install --group wx -e . ``` ### 1.3 导出语音文件 @@ -32,14 +37,66 @@ python ./WeClone-audio/get_sample_audio.py --db-path "导出数据库路径" --M ``` ## 2. 语音合成推理 +### Spark-TTS模型 +**环境安装** +可不创建新环境，直接安装依赖组到WeClone共主环境 + +```bash +uv venv .venv-sparktts --python=3.9 +source .venv-sparktts/bin/activate +uv pip install --group sparktts -e . + +cd WeClone-audio/src +git clone https://github.com/SparkAudio/Spark-TTS.git +``` + +**模型下载** + +通过python下载: +```python +from huggingface_hub import snapshot_download + +snapshot_download("SparkAudio/Spark-TTS-0.5B", local_dir="pretrained_models/Spark-TTS-0.5B") +``` + +或通过git下载: +```sh +cd WeClone-audio +mkdir -p pretrained_models + +# Make sure you have git-lfs installed (https://git-lfs.com) +git lfs install +git clone https://huggingface.co/SparkAudio/Spark-TTS-0.5B pretrained_models/Spark-TTS-0.5B +``` +使用代码推理 +```python +import os +import SparkTTS +import soundfile as sf +import torch + +from SparkTTS import SparkTTS + +model = SparkTTS("WeClone-audio/pretrained_models/Spark-TTS-0.5B", "cuda") + + +with torch.no_grad(): + wav = model.inference( + text="晚上好啊,小可爱们，该睡觉了哦", + prompt_speech_path=os.path.join(os.path.dirname(__file__), "sample.wav"), + prompt_text="对，这就是我万人敬仰的太乙真人，虽然有点婴儿肥，但也掩不住我逼人的帅气。", + ) + sf.write(os.path.join(os.path.dirname(__file__), "output.wav"), wav, samplerate=16000) +``` +### Llasa模型 ### 2.1 环境配置 ```bash -# 创建并配置推理环境 +# 创建并配置推理环境 +## 可不创建新环境，与LLaMA-Factory环境共用 uv venv .venv-xcodec --python=3.9 source .venv-xcodec/bin/activate -uv pip install -e '.[xcodec]' - +uv pip install --group xcodec -e . # 退出环境 deactivate diff --git a/WeClone-audio/pyproject.toml b/WeClone-audio/pyproject.toml deleted file mode 100644 index d0d716d..0000000 --- a/WeClone-audio/pyproject.toml +++ /dev/null @@ -1,30 +0,0 @@ -[project] -name = "WeClone-audio" -version = "0.1.0" -description = "" -authors = [ - {name = "xming521"} -] -readme = "README.md" -requires-python = ">=3.9,<3.10" -dependencies = [] - -[project.optional-dependencies] -xcodec = ["xcodec2==0.1.3"] -wx = ["pywxdump"] - -[dependency-groups] -xcodec = ["xcodec2==0.1.3"] -wx = ["pywxdump"] - -[tool.uv] -conflicts = [ - [ - { group = "xcodec" }, - { group = "wx" }, - ], -] - -[[tool.uv.index]] -url = "https://pypi.tuna.tsinghua.edu.cn/simple/" -default = true diff --git a/WeClone-audio/src/Llasa/infer.py b/WeClone-audio/src/Llasa/infer.py new file mode 100644 index 0000000..bde7a4a --- /dev/null +++ b/WeClone-audio/src/Llasa/infer.py @@ -0,0 +1,12 @@ +import os +import soundfile as sf +from text_to_speech import TextToSpeech + + +sample_audio_text = "对，这就是我万人敬仰的太乙真人，虽然有点婴儿肥，但也掩不住我逼人的帅气。" # 示例音频文本 +sample_audio_path = os.path.join(os.path.dirname(__file__), "sample.wav") # 示例音频路径 +tts = TextToSpeech(sample_audio_path, sample_audio_text) +target_text = "晚上好啊" # 生成目标文本 +result = tts.infer(target_text) +sf.write(os.path.join(os.path.dirname(__file__), "output.wav"), result[1], result[0]) # 保存生成音频 + diff --git a/WeClone-audio/src/Llasa/text_to_speech.py b/WeClone-audio/src/Llasa/text_to_speech.py new file mode 100644 index 0000000..2bb468b --- /dev/null +++ b/WeClone-audio/src/Llasa/text_to_speech.py @@ -0,0 +1,131 @@ +import os +from transformers import AutoTokenizer, AutoModelForCausalLM +import torch +import soundfile as sf +from xcodec2.modeling_xcodec2 import XCodec2Model +import torchaudio + + +class TextToSpeech: + def __init__(self, sample_audio_path, sample_audio_text): + self.sample_audio_text = sample_audio_text + # 初始化模型 + llasa_3b = "HKUSTAudio/Llasa-3B" + xcodec2 = "HKUSTAudio/xcodec2" + + self.tokenizer = AutoTokenizer.from_pretrained(llasa_3b) + self.llasa_3b_model = AutoModelForCausalLM.from_pretrained( + llasa_3b, + trust_remote_code=True, + device_map="auto", + ) + self.llasa_3b_model.eval() + + self.xcodec_model = XCodec2Model.from_pretrained(xcodec2) + self.xcodec_model.eval().cuda() + + # 处理音频 + waveform, sample_rate = torchaudio.load(sample_audio_path) + if len(waveform[0]) / sample_rate > 15: + print("已将音频裁剪至前15秒。") + waveform = waveform[:, : sample_rate * 15] + + # 检查音频是否为立体声 + if waveform.size(0) > 1: + waveform_mono = torch.mean(waveform, dim=0, keepdim=True) + else: + waveform_mono = waveform + + self.prompt_wav = torchaudio.transforms.Resample( + orig_freq=sample_rate, new_freq=16000 + )(waveform_mono) + + # Encode the prompt wav + vq_code_prompt = self.xcodec_model.encode_code(input_waveform=self.prompt_wav) + vq_code_prompt = vq_code_prompt[0, 0, :] + self.speech_ids_prefix = self.ids_to_speech_tokens(vq_code_prompt) + self.speech_end_id = self.tokenizer.convert_tokens_to_ids("<|SPEECH_GENERATION_END|>") + + def ids_to_speech_tokens(self, speech_ids): + speech_tokens_str = [] + for speech_id in speech_ids: + speech_tokens_str.append(f"<|s_{speech_id}|>") + return speech_tokens_str + + def extract_speech_ids(self, speech_tokens_str): + speech_ids = [] + for token_str in speech_tokens_str: + if token_str.startswith("<|s_") and token_str.endswith("|>"): + num_str = token_str[4:-2] + num = int(num_str) + speech_ids.append(num) + else: + print(f"Unexpected token: {token_str}") + return speech_ids + + @torch.inference_mode() + def infer(self, target_text): + if len(target_text) == 0: + return None + elif len(target_text) > 300: + print("文本过长，请保持在300字符以内。") + target_text = target_text[:300] + + input_text = self.sample_audio_text + " " + target_text + + formatted_text = ( + f"<|TEXT_UNDERSTANDING_START|>{input_text}<|TEXT_UNDERSTANDING_END|>" + ) + + chat = [ + { + "role": "user", + "content": "Convert the text to speech:" + formatted_text, + }, + { + "role": "assistant", + "content": "<|SPEECH_GENERATION_START|>" + + "".join(self.speech_ids_prefix), + }, + ] + + input_ids = self.tokenizer.apply_chat_template( + chat, tokenize=True, return_tensors="pt", continue_final_message=True + ) + input_ids = input_ids.to("cuda") + + outputs = self.llasa_3b_model.generate( + input_ids, + max_length=2048, + eos_token_id=self.speech_end_id, + do_sample=True, + top_p=1, + temperature=0.8, + ) + generated_ids = outputs[0][input_ids.shape[1] - len(self.speech_ids_prefix): -1] + + speech_tokens = self.tokenizer.batch_decode( + generated_ids, skip_special_tokens=True + ) + + speech_tokens = self.extract_speech_ids(speech_tokens) + speech_tokens = torch.tensor(speech_tokens).cuda().unsqueeze(0).unsqueeze(0) + + gen_wav = self.xcodec_model.decode_code(speech_tokens) + gen_wav = gen_wav[:, :, self.prompt_wav.shape[1]:] + + return (16000, gen_wav[0, 0, :].cpu().numpy()) + + +if __name__ == "__main__": + # 如果遇到问题，请尝试将参考音频转换为WAV或MP3格式，将其裁剪至15秒以内，并缩短提示文本。 + sample_audio_text = "对，这就是我万人敬仰的太乙真人，虽然有点婴儿肥，但也掩不住我逼人的帅气。" + sample_audio_path = os.path.join(os.path.dirname(__file__), "sample.wav") + + tts = TextToSpeech(sample_audio_path, sample_audio_text) + target_text = "晚上好啊，吃了吗您" + result = tts.infer(target_text) + sf.write(os.path.join(os.path.dirname(__file__), "output.wav"), result[1], result[0]) + target_text = "我是老北京正黄旗！" + result = tts.infer(target_text) + sf.write(os.path.join(os.path.dirname(__file__), "output1.wav"), result[1], result[0]) diff --git a/WeClone-audio/src/Spark-TTS b/WeClone-audio/src/Spark-TTS new file mode 160000 index 0000000..ee29f36 --- /dev/null +++ b/WeClone-audio/src/Spark-TTS @@ -0,0 +1 @@ +Subproject commit ee29f36806237be088e71931d9e0a4ab91b34adf diff --git a/WeClone-audio/src/SparkTTS.py b/WeClone-audio/src/SparkTTS.py new file mode 100644 index 0000000..e12bb17 --- /dev/null +++ b/WeClone-audio/src/SparkTTS.py @@ -0,0 +1,223 @@ +import re +import torch +from typing import Tuple +from pathlib import Path +from transformers import AutoTokenizer, AutoModelForCausalLM +import os +import sys +sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), "./Spark-TTS"))) +from sparktts.utils.file import load_config +from sparktts.models.audio_tokenizer import BiCodecTokenizer +from sparktts.utils.token_parser import LEVELS_MAP, GENDER_MAP, TASK_TOKEN_MAP + + +class SparkTTS: + """ + Spark-TTS for text-to-speech generation. + """ + + def __init__(self, model_dir: Path, device: torch.device = torch.device("cuda:0")): + """ + Initializes the SparkTTS model with the provided configurations and device. + + Args: + model_dir (Path): Directory containing the model and config files. + device (torch.device): The device (CPU/GPU) to run the model on. + """ + self.device = device + self.model_dir = model_dir + self.configs = load_config(f"{model_dir}/config.yaml") + self.sample_rate = self.configs["sample_rate"] + self._initialize_inference() + + def _initialize_inference(self): + """Initializes the tokenizer, model, and audio tokenizer for inference.""" + self.tokenizer = AutoTokenizer.from_pretrained(f"{self.model_dir}/LLM") + self.model = AutoModelForCausalLM.from_pretrained(f"{self.model_dir}/LLM") + self.audio_tokenizer = BiCodecTokenizer(self.model_dir, device=self.device) + self.model.to(self.device) + + def process_prompt( + self, + text: str, + prompt_speech_path: Path, + prompt_text: str = None, + ) -> Tuple[str, torch.Tensor]: + """ + Process input for voice cloning. + + Args: + text (str): The text input to be converted to speech. + prompt_speech_path (Path): Path to the audio file used as a prompt. + prompt_text (str, optional): Transcript of the prompt audio. + + Return: + Tuple[str, torch.Tensor]: Input prompt; global tokens + """ + + global_token_ids, semantic_token_ids = self.audio_tokenizer.tokenize( + prompt_speech_path + ) + global_tokens = "".join( + [f"<|bicodec_global_{i}|>" for i in global_token_ids.squeeze()] + ) + + # Prepare the input tokens for the model + if prompt_text is not None: + semantic_tokens = "".join( + [f"<|bicodec_semantic_{i}|>" for i in semantic_token_ids.squeeze()] + ) + inputs = [ + TASK_TOKEN_MAP["tts"], + "<|start_content|>", + prompt_text, + text, + "<|end_content|>", + "<|start_global_token|>", + global_tokens, + "<|end_global_token|>", + "<|start_semantic_token|>", + semantic_tokens, + ] + else: + inputs = [ + TASK_TOKEN_MAP["tts"], + "<|start_content|>", + text, + "<|end_content|>", + "<|start_global_token|>", + global_tokens, + "<|end_global_token|>", + ] + + inputs = "".join(inputs) + + return inputs, global_token_ids + + def process_prompt_control( + self, + gender: str, + pitch: str, + speed: str, + text: str, + ): + """ + Process input for voice creation. + + Args: + gender (str): female | male. + pitch (str): very_low | low | moderate | high | very_high + speed (str): very_low | low | moderate | high | very_high + text (str): The text input to be converted to speech. + + Return: + str: Input prompt + """ + assert gender in GENDER_MAP.keys() + assert pitch in LEVELS_MAP.keys() + assert speed in LEVELS_MAP.keys() + + gender_id = GENDER_MAP[gender] + pitch_level_id = LEVELS_MAP[pitch] + speed_level_id = LEVELS_MAP[speed] + + pitch_label_tokens = f"<|pitch_label_{pitch_level_id}|>" + speed_label_tokens = f"<|speed_label_{speed_level_id}|>" + gender_tokens = f"<|gender_{gender_id}|>" + + attribte_tokens = "".join( + [gender_tokens, pitch_label_tokens, speed_label_tokens] + ) + + control_tts_inputs = [ + TASK_TOKEN_MAP["controllable_tts"], + "<|start_content|>", + text, + "<|end_content|>", + "<|start_style_label|>", + attribte_tokens, + "<|end_style_label|>", + ] + + return "".join(control_tts_inputs) + + @torch.no_grad() + def inference( + self, + text: str, + prompt_speech_path: Path = None, + prompt_text: str = None, + gender: str = None, + pitch: str = None, + speed: str = None, + temperature: float = 0.8, + top_k: float = 50, + top_p: float = 0.95, + ) -> torch.Tensor: + """ + Performs inference to generate speech from text, incorporating prompt audio and/or text. + + Args: + text (str): The text input to be converted to speech. + prompt_speech_path (Path): Path to the audio file used as a prompt. + prompt_text (str, optional): Transcript of the prompt audio. + gender (str): female | male. + pitch (str): very_low | low | moderate | high | very_high + speed (str): very_low | low | moderate | high | very_high + temperature (float, optional): Sampling temperature for controlling randomness. Default is 0.8. + top_k (float, optional): Top-k sampling parameter. Default is 50. + top_p (float, optional): Top-p (nucleus) sampling parameter. Default is 0.95. + + Returns: + torch.Tensor: Generated waveform as a tensor. + """ + if gender is not None: + prompt = self.process_prompt_control(gender, pitch, speed, text) + + else: + prompt, global_token_ids = self.process_prompt( + text, prompt_speech_path, prompt_text + ) + model_inputs = self.tokenizer([prompt], return_tensors="pt").to(self.device) + + # Generate speech using the model + generated_ids = self.model.generate( + **model_inputs, + max_new_tokens=3000, + do_sample=True, + top_k=top_k, + top_p=top_p, + temperature=temperature, + ) + + # Trim the output tokens to remove the input tokens + generated_ids = [ + output_ids[len(input_ids):] + for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids) + ] + + # Decode the generated tokens into text + predicts = self.tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0] + + # Extract semantic token IDs from the generated text + pred_semantic_ids = ( + torch.tensor([int(token) for token in re.findall(r"bicodec_semantic_(\d+)", predicts)]) + .long() + .unsqueeze(0) + ) + + if gender is not None: + global_token_ids = ( + torch.tensor([int(token) for token in re.findall(r"bicodec_global_(\d+)", predicts)]) + .long() + .unsqueeze(0) + .unsqueeze(0) + ) + + # Convert semantic tokens back to waveform + wav = self.audio_tokenizer.detokenize( + global_token_ids.to(self.device).squeeze(0), + pred_semantic_ids.to(self.device), + ) + + return wav diff --git a/WeClone-audio/src/infer.py b/WeClone-audio/src/infer.py index 3da20a8..99ffcc9 100644 --- a/WeClone-audio/src/infer.py +++ b/WeClone-audio/src/infer.py @@ -1,11 +1,18 @@ import os +import SparkTTS import soundfile as sf -from text_to_speech import TextToSpeech +import torch + +from SparkTTS import SparkTTS + +model = SparkTTS("WeClone-audio/pretrained_models/Spark-TTS-0.5B", "cuda") -sample_audio_text = "对，这就是我万人敬仰的太乙真人，虽然有点婴儿肥，但也掩不住我逼人的帅气。" # 示例音频文本 -sample_audio_path = os.path.join(os.path.dirname(__file__), "sample.wav") # 示例音频路径 -tts = TextToSpeech(sample_audio_path, sample_audio_text) -target_text = "晚上好啊" # 生成目标文本 -result = tts.infer(target_text) -sf.write(os.path.join(os.path.dirname(__file__), "output.wav"), result[1], result[0]) # 保存生成音频 +with torch.no_grad(): + wav = model.inference( + text="晚上好啊,小可爱们，该睡觉了哦", + prompt_speech_path=os.path.join(os.path.dirname(__file__), "sample.wav"), + prompt_text="对，这就是我万人敬仰的太乙真人，虽然有点婴儿肥，但也掩不住我逼人的帅气。", + ) + sf.write(os.path.join(os.path.dirname(__file__), "output.wav"), wav, samplerate=16000) + print("生成成功！") diff --git a/pyproject.toml b/pyproject.toml new file mode 100644 index 0000000..cd2abd3 --- /dev/null +++ b/pyproject.toml @@ -0,0 +1,65 @@ +[project] +name = "WeClone" +version = "0.1.2" +description = "" +authors = [{ name = "xming521" }] +readme = "README.md" +requires-python = ">=3.9,<3.10" +dependencies = ["pandas", "pydantic==2.10.6"] + +# [project.optional-dependencies] +# xcodec = ["xcodec2==0.1.3"] +# wx = ["pywxdump"] +# sparktts = [ +# "einops>=0.8.1", +# "einx>=0.3.0", +# "numpy==1.26.4", +# "omegaconf>=2.3.0", +# "packaging>=24.2", +# "safetensors>=0.5.2", +# "soundfile>=0.12.1", +# "soxr>=0.5.0.post1", +# "torch>=2.5.1", +# "torchaudio>=2.5.1", +# "tqdm>=4.66.5", +# "transformers==4.45.2" +# ] + +[dependency-groups] +xcodec = ["xcodec2==0.1.3"] +wx = ["pywxdump"] +sparktts = [ + "einops>=0.8.1", + "einx>=0.3.0", + "numpy==1.26.4", + "omegaconf>=2.3.0", + "packaging>=24.2", + "safetensors>=0.5.2", + "soundfile>=0.12.1", + "soxr>=0.5.0.post1", + "torch>=2.5.1", + "torchaudio>=2.5.1", + "tqdm>=4.66.5", + "transformers==4.45.2", +] +main = ["transformers==4.45.2", "llamafactory>=0.9.2", "openai==0.28.0"] + +[tool.uv] +conflicts = [ + [ + { group = "wx" }, + { group = "sparktts" }, + ], + [ + { group = "wx" }, + { group = "main" }, + ], + [ + { group = "wx" }, + { group = "xcodec" }, + ], +] + +[[tool.uv.index]] +url = "https://pypi.tuna.tsinghua.edu.cn/simple/" +default = true diff --git a/requirements.txt b/requirements.txt index 345d3d8..db3312b 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ # LLaMA-Factory -llmtuner==0.5.3 +llmtuner # wechat itchat-uos==1.5.0.dev0 # others diff --git a/settings.json b/settings.json index f5e3b1f..725098f 100644 --- a/settings.json +++ b/settings.json @@ -46,7 +46,8 @@ "model_name_or_path": "./chatglm3-6b", "adapter_name_or_path": "./model_output", "template": "chatglm3-weclone", - "finetuning_type": "lora" + "finetuning_type": "lora", + "trust_remote_code": true }, "_comment": "adapter_name_or_path同时做为train_sft_args的output_dir " } \ No newline at end of file diff --git a/src/api_service.py b/src/api_service.py index 226e345..2f0131c 100644 --- a/src/api_service.py +++ b/src/api_service.py @@ -1,6 +1,7 @@ import os import uvicorn -from llmtuner import ChatModel, create_app +from llamafactory.chat import ChatModel +from llamafactory.api.app import create_app from template import template_register from utils.config import load_config diff --git a/src/cli_demo.py b/src/cli_demo.py index ba828f5..29dfb57 100644 --- a/src/cli_demo.py +++ b/src/cli_demo.py @@ -1,5 +1,5 @@ -from llmtuner import ChatModel -from llmtuner.extras.misc import torch_gc +from llamafactory.chat import ChatModel +from llamafactory.extras.misc import torch_gc try: diff --git a/src/evaluate.py b/src/evaluate.py index 13796c0..a09094a 100644 --- a/src/evaluate.py +++ b/src/evaluate.py @@ -1,4 +1,4 @@ -from llmtuner import Evaluator +from llamafactory.eval.evaluator import Evaluator def main(): diff --git a/src/export_model.py b/src/export_model.py index 4baeb2c..4fa5f3a 100644 --- a/src/export_model.py +++ b/src/export_model.py @@ -1,4 +1,4 @@ -from llmtuner import export_model +from llamafactory.train.tuner import export_model def main(): diff --git a/src/template.py b/src/template.py index 7b66b30..8cf3907 100644 --- a/src/template.py +++ b/src/template.py @@ -1,21 +1,24 @@ -from llmtuner.data.formatter import FunctionFormatter, StringFormatter -from llmtuner.data.template import _register_template +from llamafactory.data.formatter import FunctionFormatter, StringFormatter, ToolFormatter, EmptyFormatter +from llamafactory.data.template import register_template default_prompt = "请你扮演一名人类，不要说自己是人工智能" def template_register(): - _register_template( + register_template( name="chatglm3-weclone", default_system=( default_prompt ), format_user=StringFormatter(slots=[{"token": "<|user|>"}, "\n", "{{content}}", {"token": "<|assistant|>"}]), format_assistant=StringFormatter(slots=["\n", "{{content}}"]), - format_system=StringFormatter(slots=[{"token": "[gMASK]"}, {"token": "sop"}, {"token": "<|system|>"}, "\n", "{{content}}"]), - format_function=FunctionFormatter(slots=["{{name}}\n{{arguments}}"]), - format_observation=StringFormatter(slots=[{"token": "<|observation|>"}, "\n", "{{content}}"]), + format_system=StringFormatter(slots=[{"token": "<|system|>"}, "\n", "{{content}}"]), + format_function=FunctionFormatter(slots=["{{content}}"], tool_format="glm4"), + format_observation=StringFormatter( + slots=[{"token": "<|observation|>"}, "\n", "{{content}}", {"token": "<|assistant|>"}] + ), + format_tools=ToolFormatter(tool_format="glm4"), + format_prefix=EmptyFormatter(slots=[{"token": "[gMASK]"}, {"token": "sop"}]), stop_words=["<|user|>", "<|observation|>"], efficient_eos=True, - force_system=True ) diff --git a/src/train_pt.py b/src/train_pt.py index f57097f..4db93cd 100644 --- a/src/train_pt.py +++ b/src/train_pt.py @@ -1,4 +1,4 @@ -from llmtuner import run_exp +from llamafactory.train.tuner import run_exp from utils.config import load_config config = load_config('train_pt') diff --git a/src/train_sft.py b/src/train_sft.py index 47653bf..e3dda8b 100644 --- a/src/train_sft.py +++ b/src/train_sft.py @@ -1,4 +1,4 @@ -from llmtuner import run_exp +from llamafactory.train.tuner import run_exp from template import template_register from utils.config import load_config diff --git a/src/web_demo.py b/src/web_demo.py index 625684b..31028ed 100644 --- a/src/web_demo.py +++ b/src/web_demo.py @@ -1,4 +1,4 @@ -from llmtuner import create_web_demo +from llamafactory.webui.interface import create_web_demo from template import template_register from utils.config import load_config