Merge pull request #1758 from orbisai0security/fix/V-005-pickle-deserialization

[Security] Fix CRITICAL vulnerability: V-005

README.md

@@ -1,12 +1,12 @@
-[![SVG Banners](https://svg-banners.vercel.app/api?type=origin&text1=CosyVoice🤠&text2=Text-to-Speech%20💖%20Large%20Language%20Model&width=800&height=210)](https://github.com/Akshay090/svg-banners)
+![SVG Banners](https://svg-banners.vercel.app/api?type=origin&text1=CosyVoice🤠&text2=Text-to-Speech%20💖%20Large%20Language%20Model&width=800&height=210)
 
 ## 👉🏻 CosyVoice 👈🏻
 
-**Fun-CosyVoice 3.0**: [Demos](https://funaudiollm.github.io/cosyvoice3/); [Paper](https://arxiv.org/abs/2505.17589); [Modelscope](https://www.modelscope.cn/studios/FunAudioLLM/Fun-CosyVoice3-0.5B); [CV3-Eval](https://github.com/FunAudioLLM/CV3-Eval)
+**Fun-CosyVoice 3.0**: [Demos](https://funaudiollm.github.io/cosyvoice3/); [Paper](https://arxiv.org/pdf/2505.17589); [Modelscope](https://www.modelscope.cn/models/FunAudioLLM/Fun-CosyVoice3-0.5B-2512); [Huggingface](https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512); [CV3-Eval](https://github.com/FunAudioLLM/CV3-Eval)
 
-**CosyVoice 2.0**: [Demos](https://funaudiollm.github.io/cosyvoice2/); [Paper](https://arxiv.org/abs/2412.10117); [Modelscope](https://www.modelscope.cn/studios/iic/CosyVoice2-0.5B); [HuggingFace](https://huggingface.co/spaces/FunAudioLLM/CosyVoice2-0.5B)
+**CosyVoice 2.0**: [Demos](https://funaudiollm.github.io/cosyvoice2/); [Paper](https://arxiv.org/pdf/2412.10117); [Modelscope](https://www.modelscope.cn/models/iic/CosyVoice2-0.5B); [HuggingFace](https://huggingface.co/FunAudioLLM/CosyVoice2-0.5B)
 
-**CosyVoice 1.0**: [Demos](https://fun-audio-llm.github.io); [Paper](https://funaudiollm.github.io/pdf/CosyVoice_v1.pdf); [Modelscope](https://www.modelscope.cn/studios/iic/CosyVoice-300M)
+**CosyVoice 1.0**: [Demos](https://fun-audio-llm.github.io); [Paper](https://funaudiollm.github.io/pdf/CosyVoice_v1.pdf); [Modelscope](https://www.modelscope.cn/models/iic/CosyVoice-300M); [HuggingFace](https://huggingface.co/FunAudioLLM/CosyVoice-300M)
 
 ## Highlight🔥
 
@@ -60,23 +60,25 @@
     - [x] Fastapi server and client
 
 ## Evaluation
-| Model | Model Size | CER (%) ↓ (test-zh) | WER (%) ↓ (test-en) | CER (%) ↓ (test-hard) |
-|-------|------------|---------------------|---------------------|-----------------------|
-| Human | - | 1.26 | 2.14 | - |
-| Seed-TTS | - | 1.12 | 2.25 | 7.59 |
-| MiniMax-Speech | - | 0.83 | 1.65 | - |
-| F5-TTS | 0.3B | 1.52 | 2.00 | 8.67 |
-| SparkTTS | 0.5B | 1.20 | 1.98 | - |
-| CosyVoice2 | 0.5B | 1.45 | 2.57 | 6.83 |
-| FireRedTTS-2 | 1.5B | 1.14 | 1.95 | - |
-| IndexTTS2 | 1.5B | 1.01 | 1.52 | 7.12 |
-| VibeVoice | 1.5B | 1.16 | 3.04 | - |
-| HiggsAudio-v2 | 3B | 1.50 | 2.44 | - |
-| VoxPCM | 0.5B | 0.93 | 1.85 | 8.87 |
-| GLM-TTS | 1.5B | 1.03 | - | - |
-| GLM-TTS_RL | 1.5B | 0.89 | - | - |
-| Fun-CosyVoice3-0.5B-2512 | 0.5B | 1.21 |  2.24 | 6.71 |
-| Fun-CosyVoice3-0.5B-2512_RL | 0.5B | 0.81 | 1.68 | 5.44 |
+
+| Model | Open-Source | Model Size | test-zh<br>CER (%) ↓ | test-zh<br>SS (%) ↑ | test-en<br>WER (%) ↓ | test-en<br>SS (%) ↑ | test-hard<br>CER (%) ↓ | test-hard<br>SS (%) ↑ |
+| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
+| Human | - | - | 1.26 | 75.5 | 2.14 | 73.4 | - | - |
+| Seed-TTS | ❌ | - | 1.12 | 79.6 | 2.25 | 76.2 | 7.59 | 77.6 |
+| MiniMax-Speech | ❌ | - | 0.83 | 78.3 | 1.65 | 69.2 | - | - |
+| F5-TTS | ✅ | 0.3B | 1.52 | 74.1 | 2.00 | 64.7 | 8.67 | 71.3 |
+| Spark TTS | ✅ | 0.5B | 1.2 | 66.0 | 1.98 | 57.3 | - | - |
+| CosyVoice2 | ✅ | 0.5B | 1.45 | 75.7 | 2.57 | 65.9 | 6.83 | 72.4 |
+| FireRedTTS2 | ✅ | 1.5B | 1.14 | 73.2 | 1.95 | 66.5 | - | - |
+| Index-TTS2 | ✅ | 1.5B | 1.03 | 76.5 | 2.23 | 70.6 | 7.12 | 75.5 |
+| VibeVoice-1.5B | ✅ | 1.5B | 1.16 | 74.4 | 3.04 | 68.9 | - | - |
+| VibeVoice-Realtime | ✅ | 0.5B | - | - | 2.05 | 63.3 | - | - |
+| HiggsAudio-v2 | ✅ | 3B | 1.50 | 74.0 | 2.44 | 67.7 | - | - |
+| VoxCPM | ✅ | 0.5B | 0.93 | 77.2 | 1.85 | 72.9 | 8.87 | 73.0 |
+| GLM-TTS | ✅ | 1.5B | 1.03 | 76.1 | - | - | - | - |
+| GLM-TTS RL | ✅ | 1.5B | 0.89 | 76.4 | - | - | - | - |
+| Fun-CosyVoice3-0.5B-2512 | ✅ | 0.5B | 1.21 | 78.0 | 2.24 | 71.8 | 6.71 | 75.8 |
+| Fun-CosyVoice3-0.5B-2512_RL | ✅ | 0.5B | 0.81 | 77.4 | 1.68 | 69.5 | 5.44 | 75.0 |
 
 
 ## Install
@@ -111,7 +113,7 @@
 We strongly recommend that you download our pretrained `Fun-CosyVoice3-0.5B` `CosyVoice2-0.5B` `CosyVoice-300M` `CosyVoice-300M-SFT` `CosyVoice-300M-Instruct` model and `CosyVoice-ttsfrd` resource.
 
 ``` python
-# SDK模型下载
+# modelscope SDK model download
 from modelscope import snapshot_download
 snapshot_download('FunAudioLLM/Fun-CosyVoice3-0.5B-2512', local_dir='pretrained_models/Fun-CosyVoice3-0.5B')
 snapshot_download('iic/CosyVoice2-0.5B', local_dir='pretrained_models/CosyVoice2-0.5B')
@@ -119,6 +121,15 @@ snapshot_download('iic/CosyVoice-300M', local_dir='pretrained_models/CosyVoice-3
 snapshot_download('iic/CosyVoice-300M-SFT', local_dir='pretrained_models/CosyVoice-300M-SFT')
 snapshot_download('iic/CosyVoice-300M-Instruct', local_dir='pretrained_models/CosyVoice-300M-Instruct')
 snapshot_download('iic/CosyVoice-ttsfrd', local_dir='pretrained_models/CosyVoice-ttsfrd')
+
+# for oversea users, huggingface SDK model download
+from huggingface_hub import snapshot_download
+snapshot_download('FunAudioLLM/Fun-CosyVoice3-0.5B-2512', local_dir='pretrained_models/Fun-CosyVoice3-0.5B')
+snapshot_download('FunAudioLLM/CosyVoice2-0.5B', local_dir='pretrained_models/CosyVoice2-0.5B')
+snapshot_download('FunAudioLLM/CosyVoice-300M', local_dir='pretrained_models/CosyVoice-300M')
+snapshot_download('FunAudioLLM/CosyVoice-300M-SFT', local_dir='pretrained_models/CosyVoice-300M-SFT')
+snapshot_download('FunAudioLLM/CosyVoice-300M-Instruct', local_dir='pretrained_models/CosyVoice-300M-Instruct')
+snapshot_download('FunAudioLLM/CosyVoice-ttsfrd', local_dir='pretrained_models/CosyVoice-ttsfrd')
 ```
 
 Optionally, you can unzip `ttsfrd` resource and install `ttsfrd` package for better text normalization performance.
@@ -140,15 +151,19 @@ Follow the code in `example.py` for detailed usage of each model.
 python example.py
 ```
 
-#### CosyVoice2 vllm Usage
-If you want to use vllm for inference, please install `vllm==v0.9.0`. Older vllm version do not support CosyVoice2 inference.
+#### vLLM Usage
+CosyVoice2/3 now supports **vLLM 0.11.x+ (V1 engine)** and **vLLM 0.9.0 (legacy)**.
+Older vllm version(<0.9.0) do not support CosyVoice inference, and versions in between (e.g., 0.10.x) are not tested.
 
-Notice that `vllm==v0.9.0` has a lot of specific requirements, for example `torch==2.7.0`. You can create a new env to in case your hardward do not support vllm and old env is corrupted.
+Notice that `vllm` has a lot of specific requirements. You can create a new env to in case your hardward do not support vllm and old env is corrupted.
 
 ``` sh
 conda create -n cosyvoice_vllm --clone cosyvoice
 conda activate cosyvoice_vllm
-pip install vllm==v0.9.0 transformers==4.51.3 -i https://mirrors.aliyun.com/pypi/simple/ --trusted-host=mirrors.aliyun.com
+# for vllm==0.9.0
+pip install vllm==v0.9.0 transformers==4.51.3 numpy==1.26.4 -i https://mirrors.aliyun.com/pypi/simple/ --trusted-host=mirrors.aliyun.com
+# for vllm>=0.11.0
+pip install vllm==v0.11.0 transformers==4.57.1 numpy==1.26.4 -i https://mirrors.aliyun.com/pypi/simple/ --trusted-host=mirrors.aliyun.com
 python vllm_example.py
 ```
 
@@ -165,7 +180,7 @@ python3 webui.py --port 50000 --model_dir pretrained_models/CosyVoice-300M
 
 #### Advanced Usage
 
-For advanced users, we have provided training and inference scripts in `examples/libritts/cosyvoice/run.sh`.
+For advanced users, we have provided training and inference scripts in `examples/libritts`.
 
 #### Build for deployment
 

cosyvoice/bin/export_jit.py

@@ -24,9 +24,7 @@ ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
 sys.path.append('{}/../..'.format(ROOT_DIR))
 sys.path.append('{}/../../third_party/Matcha-TTS'.format(ROOT_DIR))
 from cosyvoice.cli.cosyvoice import AutoModel
-from cosyvoice.cli.model import CosyVoiceModel, CosyVoice2Model, CosyVoice3Model
 from cosyvoice.utils.file_utils import logging
-from cosyvoice.utils.class_utils import get_model_type
 
 
 def get_args():
@@ -61,15 +59,7 @@ def main():
 
     model = AutoModel(model_dir=args.model_dir)
 
-    if get_model_type(model.model) == CosyVoiceModel:
-        # 1. export flow encoder
-        flow_encoder = model.model.flow.encoder
-        script = get_optimized_script(flow_encoder)
-        script.save('{}/flow.encoder.fp32.zip'.format(args.model_dir))
-        script = get_optimized_script(flow_encoder.half())
-        script.save('{}/flow.encoder.fp16.zip'.format(args.model_dir))
-        logging.info('successfully export flow_encoder')
-    elif get_model_type(model.model) == CosyVoice2Model:
+    if model.__class__.__name__ == 'CosyVoice':
         # 1. export llm text_encoder
         llm_text_encoder = model.model.llm.text_encoder
         script = get_optimized_script(llm_text_encoder)
@@ -93,6 +83,14 @@ def main():
         script = get_optimized_script(flow_encoder.half())
         script.save('{}/flow.encoder.fp16.zip'.format(args.model_dir))
         logging.info('successfully export flow_encoder')
+    elif model.__class__.__name__ == 'CosyVoice2':
+        # 1. export flow encoder
+        flow_encoder = model.model.flow.encoder
+        script = get_optimized_script(flow_encoder)
+        script.save('{}/flow.encoder.fp32.zip'.format(args.model_dir))
+        script = get_optimized_script(flow_encoder.half())
+        script.save('{}/flow.encoder.fp16.zip'.format(args.model_dir))
+        logging.info('successfully export flow_encoder')
     else:
         raise ValueError('unsupported model type')
 

cosyvoice/cli/cosyvoice.py

@@ -89,6 +89,8 @@ class CosyVoice:
                 start_time = time.time()
 
     def inference_zero_shot(self, tts_text, prompt_text, prompt_wav, zero_shot_spk_id='', stream=False, speed=1.0, text_frontend=True):
+        if self.__class__.__name__ == 'CosyVoice3' and '<|endofprompt|>' not in prompt_text + tts_text:
+            logging.warning('<|endofprompt|> not found in CosyVoice3 inference, check your input text')
         prompt_text = self.frontend.text_normalize(prompt_text, split=False, text_frontend=text_frontend)
         for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)):
             if (not isinstance(i, Generator)) and len(i) < 0.5 * len(prompt_text):
@@ -114,7 +116,7 @@ class CosyVoice:
                 start_time = time.time()
 
     def inference_instruct(self, tts_text, spk_id, instruct_text, stream=False, speed=1.0, text_frontend=True):
-        assert isinstance(self.model, CosyVoiceModel), 'inference_instruct is only implemented for CosyVoice!'
+        assert self.__class__.__name__ == 'CosyVoice', 'inference_instruct is only implemented for CosyVoice!'
         instruct_text = self.frontend.text_normalize(instruct_text, split=False, text_frontend=text_frontend)
         for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)):
             model_input = self.frontend.frontend_instruct(i, spk_id, instruct_text)

cosyvoice/cli/frontend.py

@@ -20,18 +20,9 @@ import numpy as np
 import whisper
 from typing import Callable
 import torchaudio.compliance.kaldi as kaldi
-import torchaudio
 import os
 import re
 import inflect
-try:
-    import ttsfrd
-    use_ttsfrd = True
-except ImportError:
-    print("failed to import ttsfrd, use wetext instead")
-    from wetext import Normalizer as ZhNormalizer
-    from wetext import Normalizer as EnNormalizer
-    use_ttsfrd = False
 from cosyvoice.utils.file_utils import logging, load_wav
 from cosyvoice.utils.frontend_utils import contains_chinese, replace_blank, replace_corner_mark, remove_bracket, spell_out_number, split_paragraph, is_only_punctuation
 
@@ -56,21 +47,33 @@ class CosyVoiceFrontEnd:
                                                                      providers=["CUDAExecutionProvider" if torch.cuda.is_available() else
                                                                                 "CPUExecutionProvider"])
         if os.path.exists(spk2info):
-            self.spk2info = torch.load(spk2info, map_location=self.device)
+            self.spk2info = torch.load(spk2info, map_location=self.device, weights_only=True)
         else:
             self.spk2info = {}
         self.allowed_special = allowed_special
-        self.use_ttsfrd = use_ttsfrd
-        if self.use_ttsfrd:
+        self.inflect_parser = inflect.engine()
+        # NOTE compatible when no text frontend tool is avaliable
+        try:
+            import ttsfrd
             self.frd = ttsfrd.TtsFrontendEngine()
             ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
             assert self.frd.initialize('{}/../../pretrained_models/CosyVoice-ttsfrd/resource'.format(ROOT_DIR)) is True, \
                 'failed to initialize ttsfrd resource'
             self.frd.set_lang_type('pinyinvg')
-        else:
-            self.zh_tn_model = ZhNormalizer(remove_erhua=False)
-            self.en_tn_model = EnNormalizer()
-            self.inflect_parser = inflect.engine()
+            self.text_frontend = 'ttsfrd'
+            logging.info('use ttsfrd frontend')
+        except:
+            try:
+                from wetext import Normalizer as ZhNormalizer
+                from wetext import Normalizer as EnNormalizer
+                self.zh_tn_model = ZhNormalizer(remove_erhua=False)
+                self.en_tn_model = EnNormalizer()
+                self.text_frontend = 'wetext'
+                logging.info('use wetext frontend')
+            except:
+                self.text_frontend = ''
+                logging.info('no frontend is avaliable')
+
 
     def _extract_text_token(self, text):
         if isinstance(text, Generator):
@@ -131,12 +134,13 @@ class CosyVoiceFrontEnd:
         if text_frontend is False or text == '':
             return [text] if split is True else text
         text = text.strip()
-        if self.use_ttsfrd:
+        if self.text_frontend == 'ttsfrd':
             texts = [i["text"] for i in json.loads(self.frd.do_voicegen_frd(text))["sentences"]]
             text = ''.join(texts)
         else:
             if contains_chinese(text):
-                text = self.zh_tn_model.normalize(text)
+                if self.text_frontend == 'wetext':
+                    text = self.zh_tn_model.normalize(text)
                 text = text.replace("\n", "")
                 text = replace_blank(text)
                 text = replace_corner_mark(text)
@@ -147,7 +151,8 @@ class CosyVoiceFrontEnd:
                 texts = list(split_paragraph(text, partial(self.tokenizer.encode, allowed_special=self.allowed_special), "zh", token_max_n=80,
                                              token_min_n=60, merge_len=20, comma_split=False))
             else:
-                text = self.en_tn_model.normalize(text)
+                if self.text_frontend == 'wetext':
+                    text = self.en_tn_model.normalize(text)
                 text = spell_out_number(text, self.inflect_parser)
                 texts = list(split_paragraph(text, partial(self.tokenizer.encode, allowed_special=self.allowed_special), "en", token_max_n=80,
                                              token_min_n=60, merge_len=20, comma_split=False))
@@ -178,7 +183,7 @@ class CosyVoiceFrontEnd:
                            'prompt_speech_feat': speech_feat, 'prompt_speech_feat_len': speech_feat_len,
                            'llm_embedding': embedding, 'flow_embedding': embedding}
         else:
-            model_input = self.spk2info[zero_shot_spk_id]
+            model_input = {**self.spk2info[zero_shot_spk_id]}
         model_input['text'] = tts_text_token
         model_input['text_len'] = tts_text_token_len
         return model_input

cosyvoice/cli/model.py

@@ -60,14 +60,15 @@ class CosyVoiceModel:
         self.mel_overlap_dict = {}
         self.flow_cache_dict = {}
         self.hift_cache_dict = {}
+        self.silent_tokens = []
 
     def load(self, llm_model, flow_model, hift_model):
-        self.llm.load_state_dict(torch.load(llm_model, map_location=self.device), strict=True)
+        self.llm.load_state_dict(torch.load(llm_model, map_location=self.device, weights_only=True), strict=True)
         self.llm.to(self.device).eval()
-        self.flow.load_state_dict(torch.load(flow_model, map_location=self.device), strict=True)
+        self.flow.load_state_dict(torch.load(flow_model, map_location=self.device, weights_only=True), strict=True)
         self.flow.to(self.device).eval()
         # in case hift_model is a hifigan model
-        hift_state_dict = {k.replace('generator.', ''): v for k, v in torch.load(hift_model, map_location=self.device).items()}
+        hift_state_dict = {k.replace('generator.', ''): v for k, v in torch.load(hift_model, map_location=self.device, weights_only=True).items()}
         self.hift.load_state_dict(hift_state_dict, strict=True)
         self.hift.to(self.device).eval()
 
@@ -98,26 +99,33 @@ class CosyVoiceModel:
         return {'min_shape': min_shape, 'opt_shape': opt_shape, 'max_shape': max_shape, 'input_names': input_names}
 
     def llm_job(self, text, prompt_text, llm_prompt_speech_token, llm_embedding, uuid):
+        cur_silent_token_num, max_silent_token_num = 0, 5
         with self.llm_context, torch.cuda.amp.autocast(self.fp16 is True and hasattr(self.llm, 'vllm') is False):
             if isinstance(text, Generator):
-                assert isinstance(self, CosyVoice2Model) and not hasattr(self.llm, 'vllm'), 'streaming input text is only implemented for CosyVoice2 and do not support vllm!'
-                for i in self.llm.inference_bistream(text=text,
+                assert (self.__class__.__name__ != 'CosyVoiceModel') and not hasattr(self.llm, 'vllm'), 'streaming input text is only implemented for CosyVoice2/3 and do not support vllm!'
+                token_generator = self.llm.inference_bistream(text=text,
+                                                              prompt_text=prompt_text.to(self.device),
+                                                              prompt_text_len=torch.tensor([prompt_text.shape[1]], dtype=torch.int32).to(self.device),
+                                                              prompt_speech_token=llm_prompt_speech_token.to(self.device),
+                                                              prompt_speech_token_len=torch.tensor([llm_prompt_speech_token.shape[1]], dtype=torch.int32).to(self.device),
+                                                              embedding=llm_embedding.to(self.device))
+            else:
+                token_generator = self.llm.inference(text=text.to(self.device),
+                                                     text_len=torch.tensor([text.shape[1]], dtype=torch.int32).to(self.device),
                                                      prompt_text=prompt_text.to(self.device),
                                                      prompt_text_len=torch.tensor([prompt_text.shape[1]], dtype=torch.int32).to(self.device),
                                                      prompt_speech_token=llm_prompt_speech_token.to(self.device),
                                                      prompt_speech_token_len=torch.tensor([llm_prompt_speech_token.shape[1]], dtype=torch.int32).to(self.device),
-                                                     embedding=llm_embedding.to(self.device)):
-                    self.tts_speech_token_dict[uuid].append(i)
-            else:
-                for i in self.llm.inference(text=text.to(self.device),
-                                            text_len=torch.tensor([text.shape[1]], dtype=torch.int32).to(self.device),
-                                            prompt_text=prompt_text.to(self.device),
-                                            prompt_text_len=torch.tensor([prompt_text.shape[1]], dtype=torch.int32).to(self.device),
-                                            prompt_speech_token=llm_prompt_speech_token.to(self.device),
-                                            prompt_speech_token_len=torch.tensor([llm_prompt_speech_token.shape[1]], dtype=torch.int32).to(self.device),
-                                            embedding=llm_embedding.to(self.device),
-                                            uuid=uuid):
-                    self.tts_speech_token_dict[uuid].append(i)
+                                                     embedding=llm_embedding.to(self.device),
+                                                     uuid=uuid)  
+            for i in token_generator:
+                if i in self.silent_tokens:
+                    cur_silent_token_num += 1
+                    if cur_silent_token_num > max_silent_token_num:
+                        continue
+                else:
+                    cur_silent_token_num = 0
+                self.tts_speech_token_dict[uuid].append(i)
         self.llm_end_dict[uuid] = True
 
     def vc_job(self, source_speech_token, uuid):
@@ -260,6 +268,7 @@ class CosyVoice2Model(CosyVoiceModel):
         self.tts_speech_token_dict = {}
         self.llm_end_dict = {}
         self.hift_cache_dict = {}
+        self.silent_tokens = []
 
     def load_jit(self, flow_encoder_model):
         flow_encoder = torch.jit.load(flow_encoder_model, map_location=self.device)
@@ -401,6 +410,8 @@ class CosyVoice3Model(CosyVoice2Model):
         self.tts_speech_token_dict = {}
         self.llm_end_dict = {}
         self.hift_cache_dict = {}
+        # FSQ silent and breath token
+        self.silent_tokens = [1, 2, 28, 29, 55, 248, 494, 2241, 2242, 2322, 2323]
 
     def token2wav(self, token, prompt_token, prompt_feat, embedding, token_offset, uuid, stream=False, finalize=False, speed=1.0):
         with torch.cuda.amp.autocast(self.fp16):

cosyvoice/dataset/dataset.py

@@ -145,7 +145,11 @@ def Dataset(data_list_file,
                        shuffle=shuffle,
                        partition=partition)
     # map partial arg to padding func
-    data_pipeline[-1] = partial(data_pipeline[-1], gan=gan, dpo=dpo)
+    for i in range(1, len(data_pipeline)):
+        if data_pipeline[i].func.__name__ == 'compute_fbank':
+            data_pipeline[i] = partial(data_pipeline[i], token_mel_ratio=0)
+        if data_pipeline[i].func.__name__ == 'padding':
+            data_pipeline[i] = partial(data_pipeline[i], gan=gan, dpo=dpo)
     for func in data_pipeline:
         dataset = Processor(dataset, func, mode=mode)
     return dataset

cosyvoice/dataset/processor.py

@@ -26,7 +26,7 @@ import pyworld as pw
 AUDIO_FORMAT_SETS = {'flac', 'mp3', 'm4a', 'ogg', 'opus', 'wav', 'wma'}
 
 
-def parquet_opener(data, mode='train', tts_data={}):
+def parquet_opener(data, mode='train'):
     """ Give url or local file, return file descriptor
         Inplace operation.
 
@@ -44,12 +44,8 @@ def parquet_opener(data, mode='train', tts_data={}):
                 df = df.to_pandas()
                 for i in range(len(df)):
                     sample.update(dict(df.loc[i]))
-                    if mode == 'train':
-                        # NOTE do not return sample directly, must initialize a new dict
-                        yield {**sample}
-                    else:
-                        for index, text in enumerate(tts_data[df.loc[i, 'utt']]):
-                            yield {**sample, 'tts_index': index, 'tts_text': text}
+                    # NOTE do not return sample directly, must initialize a new dict
+                    yield {**sample}
         except Exception as ex:
             logging.warning('Failed to open {}, ex info {}'.format(url, ex))
 

cosyvoice/flow/flow.py

@@ -332,8 +332,9 @@ class CausalMaskedDiffWithDiT(torch.nn.Module):
         token = self.input_embedding(torch.clamp(token, min=0)) * mask
 
         # text encode
-        h, h_lengths = self.encoder(token, token_len, streaming=streaming)
-        h = self.encoder_proj(h)
+        h = self.pre_lookahead_layer(token)
+        h = h.repeat_interleave(self.token_mel_ratio, dim=1)
+        mask = mask.repeat_interleave(self.token_mel_ratio, dim=1).squeeze(dim=-1)
 
         # get conditions
         conds = torch.zeros(feat.shape, device=token.device)
@@ -344,7 +345,6 @@ class CausalMaskedDiffWithDiT(torch.nn.Module):
             conds[i, :index] = feat[i, :index]
         conds = conds.transpose(1, 2)
 
-        mask = (~make_pad_mask(h_lengths.sum(dim=-1).squeeze(dim=1))).to(h)
         loss, _ = self.decoder.compute_loss(
             feat.transpose(1, 2).contiguous(),
             mask.unsqueeze(1),

cosyvoice/llm/llm.py

@@ -301,18 +301,23 @@ class Qwen2LM(TransformerLM):
         self.stop_token_ids = [speech_token_size + i for i in range(3)]
         self.vllm_output_queue = {}
 
-    def prepare_lm_input_target(self, sos_emb, text_token, text_token_emb, text_token_len, task_id_emb, speech_token, speech_token_emb, speech_token_len):
+    def prepare_lm_input_target(self, sos_emb, text_token, text_token_emb, text_token_len, task_id_emb, speech_token, speech_token_emb, speech_token_len, instruct_token=None, instruct_token_emb=None, instruct_token_len=None):
         lm_target, lm_input = [], []
         text_token = unpad_sequence(text_token, text_token_len.cpu(), batch_first=True)
         speech_token = unpad_sequence(speech_token, speech_token_len.cpu(), batch_first=True)
         text_token_emb = unpad_sequence(text_token_emb, text_token_len.cpu(), batch_first=True)
         speech_token_emb = unpad_sequence(speech_token_emb, speech_token_len.cpu(), batch_first=True)
+        # NOTE add instruct_token in CosyVoice3
+        if instruct_token is not None and instruct_token_emb is not None and instruct_token_len is not None:
+            instruct_token = unpad_sequence(instruct_token, instruct_token_len.cpu(), batch_first=True)
+            instruct_token_emb = unpad_sequence(instruct_token_emb, instruct_token_len.cpu(), batch_first=True)
         for i in range(len(text_token)):
             # bistream sequence
             if random.random() < 0.5 and speech_token_len[i] / text_token_len[i] > self.mix_ratio[1] / self.mix_ratio[0]:
-                this_lm_target, this_lm_input = [], []
-                this_lm_target.append(IGNORE_ID)
-                this_lm_input.append(sos_emb.squeeze(dim=0))
+                this_lm_target, this_lm_input = [IGNORE_ID], [sos_emb.squeeze(dim=0)]
+                if instruct_token is not None and instruct_token_emb is not None and instruct_token_len is not None:
+                    this_lm_target += [IGNORE_ID] * instruct_token_len[i]
+                    this_lm_input.append(instruct_token_emb[i])
                 for j in range(((text_token_len[i] + 1) / self.mix_ratio[0]).ceil().int().item()):
                     this_text_token = text_token[i][j * self.mix_ratio[0]: (j + 1) * self.mix_ratio[0]].tolist()
                     this_speech_token = speech_token[i][j * self.mix_ratio[1]: (j + 1) * self.mix_ratio[1]].tolist()
@@ -333,8 +338,8 @@ class Qwen2LM(TransformerLM):
                 this_lm_target, this_lm_input = torch.tensor(this_lm_target), torch.concat(this_lm_input, dim=0)
             # unistream sequence
             else:
-                this_lm_target = torch.tensor([IGNORE_ID] * (1 + text_token_len[i]) + speech_token[i].tolist() + [self.eos_token])
-                this_lm_input = torch.concat([sos_emb.squeeze(dim=0), text_token_emb[i], task_id_emb.squeeze(dim=0), speech_token_emb[i]], dim=0)
+                this_lm_target = torch.tensor([IGNORE_ID] * (1 + instruct_token_len[i] + text_token_len[i]) + speech_token[i].tolist() + [self.eos_token])
+                this_lm_input = torch.concat([sos_emb.squeeze(dim=0), instruct_token_emb[i], text_token_emb[i], task_id_emb.squeeze(dim=0), speech_token_emb[i]], dim=0)
             lm_target.append(this_lm_target)
             lm_input.append(this_lm_input)
         lm_input_len = torch.tensor([i.size(0) for i in lm_input], dtype=torch.int32)
@@ -681,6 +686,7 @@ class CosyVoice3LM(Qwen2LM):
 
         # 1. encode text_token
         text_token_emb = self.llm.model.model.embed_tokens(text_token)
+        instruct_token_emb = self.llm.model.model.embed_tokens(instruct_token)
 
         # 3. sos and task_id
         sos_emb = self.speech_embedding.weight[self.sos].reshape(1, 1, -1)
@@ -691,14 +697,14 @@ class CosyVoice3LM(Qwen2LM):
 
         # 3. prepare llm_input/target
         lm_target, lm_input, lm_input_len = self.prepare_lm_input_target(sos_emb, text_token, text_token_emb, text_token_len, task_id_emb,
-                                                                         speech_token, speech_token_emb, speech_token_len)
+                                                                         speech_token, speech_token_emb, speech_token_len, instruct_token, instruct_token_emb, instruct_token_len)
         lm_target = lm_target.to(device)
 
         # 4. run lm forward
         lm_output, lm_output_mask = self.llm(lm_input, lm_input_len.to(device))
         logits = self.llm_decoder(lm_output)
         loss = self.criterion_ce(logits, lm_target.to(device))
-        acc = th_accuracy(logits.view(-1, self.speech_token_size + 3), lm_target, ignore_label=IGNORE_ID)
+        acc = th_accuracy(logits.view(-1, self.speech_token_size + 200), lm_target, ignore_label=IGNORE_ID)
         return {'loss': loss, 'acc': acc}
 
     @torch.inference_mode()

cosyvoice/utils/train_utils.py

@@ -53,7 +53,7 @@ def init_distributed(args):
 def init_dataset_and_dataloader(args, configs, gan, dpo):
     data_pipeline = configs['data_pipeline_gan'] if gan is True else configs['data_pipeline']
     train_dataset = Dataset(args.train_data, data_pipeline=data_pipeline, mode='train', gan=gan, dpo=dpo, shuffle=True, partition=True)
-    cv_dataset = Dataset(args.cv_data, data_pipeline=data_pipeline, mode='train', gan=gan, dpo=dpo, shuffle=False, partition=False)
+    cv_dataset = Dataset(args.cv_data, data_pipeline=data_pipeline, mode='dev', gan=gan, dpo=dpo, shuffle=False, partition=False)
 
     # do not use persistent_workers=True, as whisper tokenizer opens tiktoken file each time when the for loop starts
     train_data_loader = DataLoader(train_dataset,
@@ -164,18 +164,18 @@ def init_optimizer_and_scheduler(args, configs, model, gan):
             raise ValueError("unknown scheduler: " + configs['train_conf'])
 
         if configs['train_conf']['optim_d'] == 'adam':
-            optimizer_d = optim.Adam(model.module.discriminator.parameters(), **configs['train_conf']['optim_conf'])
+            optimizer_d = optim.Adam(model.module.discriminator.parameters(), **configs['train_conf']['optim_conf_d'])
         elif configs['train_conf']['optim_d'] == 'adamw':
-            optimizer_d = optim.AdamW(model.module.discriminator.parameters(), **configs['train_conf']['optim_conf'])
+            optimizer_d = optim.AdamW(model.module.discriminator.parameters(), **configs['train_conf']['optim_conf_d'])
         else:
             raise ValueError("unknown optimizer: " + configs['train_conf'])
 
         if configs['train_conf']['scheduler_d'] == 'warmuplr':
             scheduler_type = WarmupLR
-            scheduler_d = WarmupLR(optimizer_d, **configs['train_conf']['scheduler_conf'])
+            scheduler_d = WarmupLR(optimizer_d, **configs['train_conf']['scheduler_d'])
         elif configs['train_conf']['scheduler_d'] == 'NoamHoldAnnealing':
             scheduler_type = NoamHoldAnnealing
-            scheduler_d = NoamHoldAnnealing(optimizer_d, **configs['train_conf']['scheduler_conf'])
+            scheduler_d = NoamHoldAnnealing(optimizer_d, **configs['train_conf']['scheduler_d'])
         elif configs['train_conf']['scheduler'] == 'constantlr':
             scheduler_type = ConstantLR
             scheduler_d = ConstantLR(optimizer_d)

cosyvoice/vllm/cosyvoice2.py

@@ -23,6 +23,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Inference-only Qwen2 model compatible with HuggingFace weights."""
+from typing import Optional
+from packaging.version import parse as vparse
+import vllm
+
+# vLLM-0.11.0+ only support V1 engine
+VLLM_V1_ENGINE_ONLY: bool = vparse(vllm.__version__) >= vparse("0.11.0")
+if VLLM_V1_ENGINE_ONLY:
+    from vllm.v1.sample.metadata import SamplingMetadata
+
 from vllm.model_executor.models.qwen2 import *
 
 
@@ -87,10 +96,14 @@ class CosyVoice2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
     def compute_logits(
         self,
         hidden_states: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
+        sampling_metadata: Optional[SamplingMetadata] = None,
     ) -> Optional[torch.Tensor]:
-        logits = self.logits_processor(self.lm_head, hidden_states,
-                                       sampling_metadata, self.lm_head.bias)
+        if VLLM_V1_ENGINE_ONLY:
+            logits = self.logits_processor(self.lm_head, hidden_states, 
+                                           self.lm_head.bias)
+        else:
+            logits = self.logits_processor(self.lm_head, hidden_states,
+                                           sampling_metadata, self.lm_head.bias)
         return logits
 
     def load_weights(self, weights: Iterable[tuple[str,

docker/Dockerfile

@@ -4,7 +4,7 @@ ARG VENV_NAME="cosyvoice"
 ENV VENV=$VENV_NAME
 ENV LANG=C.UTF-8 LC_ALL=C.UTF-8
 
-ENV DEBIAN_FRONTEN=noninteractive
+ENV DEBIAN_FRONTEND=noninteractive
 ENV PYTHONUNBUFFERED=1
 SHELL ["/bin/bash", "--login", "-c"]
 

example.py

@@ -18,7 +18,7 @@ def cosyvoice_example():
     # zero_shot usage
     for i, j in enumerate(cosyvoice.inference_zero_shot('收到好友从远方寄来的生日礼物，那份意外的惊喜与深深的祝福让我心中充满了甜蜜的快乐，笑容如花儿般绽放。', '希望你以后能够做的比我还好呦。', './asset/zero_shot_prompt.wav')):
         torchaudio.save('zero_shot_{}.wav'.format(i), j['tts_speech'], cosyvoice.sample_rate)
-    # cross_lingual usage, <|zh|><|en|><|jp|><|yue|><|ko|> for Chinese/English/Japanese/Cantonese/Korean
+    # cross_lingual usage, <|zh|><|en|><|ja|><|yue|><|ko|> for Chinese/English/Japanese/Cantonese/Korean
     for i, j in enumerate(cosyvoice.inference_cross_lingual('<|en|>And then later on, fully acquiring that company. So keeping management in line, interest in line with the asset that\'s coming into the family is a reason why sometimes we don\'t buy the whole thing.',
                                                             './asset/cross_lingual_prompt.wav')):
         torchaudio.save('cross_lingual_{}.wav'.format(i), j['tts_speech'], cosyvoice.sample_rate)

examples/libritts/cosyvoice/local/prepare_data.py

@@ -40,11 +40,10 @@ def main():
     with open('{}/spk2utt'.format(args.des_dir), 'w') as f:
         for k, v in spk2utt.items():
             f.write('{} {}\n'.format(k, ' '.join(v)))
-    if args.instruct is True:
+    if args.instruct != '':
         with open('{}/instruct'.format(args.des_dir), 'w') as f:
             for k, v in utt2text.items():
-                # NOTE in CosyVoice3, we add instruct in sequence
-                f.write('{} You are a helpful assistant.<|endofprompt|>\n'.format(k, v))
+                f.write('{} {}\n'.format(k, args.instruct))
     return
 
 
@@ -55,8 +54,6 @@ if __name__ == "__main__":
     parser.add_argument('--des_dir',
                         type=str)
     parser.add_argument('--instruct',
-                        action='store_true',
-                        default=False,
-                        help='create instruct file or not')
+                        type=str)
     args = parser.parse_args()
     main()

examples/libritts/cosyvoice2/run.sh

@@ -66,7 +66,6 @@ if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
   fi
   cat data/{train-clean-100,train-clean-360,train-other-500}/parquet/data.list > data/train.data.list
   cat data/{dev-clean,dev-other}/parquet/data.list > data/dev.data.list
-  # NOTE will update llm/hift training later
   for model in llm flow hifigan; do
     torchrun --nnodes=1 --nproc_per_node=$num_gpus \
         --rdzv_id=$job_id --rdzv_backend="c10d" --rdzv_endpoint="localhost:1234" \

examples/libritts/cosyvoice3/conf/cosyvoice3.yaml

@@ -20,7 +20,7 @@ num_decoding_left_chunks: -1 # streaming inference flow decoder left chunk size,
 # model params
 # for all class/function included in this repo, we use !<name> or !<new> for intialization, so that user may find all corresponding class/function according to one single yaml.
 # for system/third_party class/function, we do not require this.
-llm: !new:cosyvoice.llm.llm.Qwen2LM
+llm: !new:cosyvoice.llm.llm.CosyVoice3LM
     llm_input_size: !ref <llm_input_size>
     llm_output_size: !ref <llm_output_size>
     speech_token_size: 6561
@@ -35,8 +35,8 @@ llm: !new:cosyvoice.llm.llm.Qwen2LM
         win_size: 10
         tau_r: 0.1
 
-flow: !new:cosyvoice.flow.flow.CausalMaskedDiffWithXvec
-    input_size: 512
+flow: !new:cosyvoice.flow.flow.CausalMaskedDiffWithDiT
+    input_size: 80
     output_size: 80
     spk_embed_dim: !ref <spk_embed_dim>
     output_type: 'mel'
@@ -45,22 +45,10 @@ flow: !new:cosyvoice.flow.flow.CausalMaskedDiffWithXvec
     only_mask_loss: True
     token_mel_ratio: !ref <token_mel_ratio>
     pre_lookahead_len: 3
-    encoder: !new:cosyvoice.transformer.upsample_encoder.UpsampleConformerEncoder
-        output_size: 512
-        attention_heads: 8
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.1
-        normalize_before: True
-        input_layer: 'linear'
-        pos_enc_layer_type: 'rel_pos_espnet'
-        selfattention_layer_type: 'rel_selfattn'
-        input_size: 512
-        use_cnn_module: False
-        macaron_style: False
-        static_chunk_size: !ref <chunk_size>
+    pre_lookahead_layer: !new:cosyvoice.transformer.upsample_encoder.PreLookaheadLayer
+        in_channels: 80
+        channels: 1024
+        pre_lookahead_len: 3
     decoder: !new:cosyvoice.flow.flow_matching.CausalConditionalCFM
         in_channels: 240
         n_spks: 1
@@ -73,20 +61,20 @@ flow: !new:cosyvoice.flow.flow.CausalMaskedDiffWithXvec
                 training_cfg_rate: 0.2
                 inference_cfg_rate: 0.7
                 reg_loss_type: 'l1'
-        estimator: !new:cosyvoice.flow.decoder.CausalConditionalDecoder
-            in_channels: 320
+        estimator: !new:cosyvoice.flow.DiT.dit.DiT
+            dim: 1024
+            depth: 22
+            heads: 16
+            dim_head: 64
+            ff_mult: 2
+            mel_dim: 80
+            mu_dim: 80
+            spk_dim: 80
             out_channels: 80
-            channels: [256]
-            dropout: 0.0
-            attention_head_dim: 64
-            n_blocks: 4
-            num_mid_blocks: 12
-            num_heads: 8
-            act_fn: 'gelu'
             static_chunk_size: !ref <chunk_size> * <token_mel_ratio>
             num_decoding_left_chunks: !ref <num_decoding_left_chunks>
 
-hift: !new:cosyvoice.hifigan.generator.HiFTGenerator
+hift: !new:cosyvoice.hifigan.generator.CausalHiFTGenerator
     in_channels: 80
     base_channels: 512
     nb_harmonics: 8
@@ -105,7 +93,8 @@ hift: !new:cosyvoice.hifigan.generator.HiFTGenerator
     source_resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
     lrelu_slope: 0.1
     audio_limit: 0.99
-    f0_predictor: !new:cosyvoice.hifigan.f0_predictor.ConvRNNF0Predictor
+    conv_pre_look_right: 4
+    f0_predictor: !new:cosyvoice.hifigan.f0_predictor.CausalConvRNNF0Predictor
         num_class: 1
         in_channels: 80
         cond_channels: 512
@@ -134,6 +123,7 @@ parquet_opener: !name:cosyvoice.dataset.processor.parquet_opener
 get_tokenizer: !name:cosyvoice.tokenizer.tokenizer.get_qwen_tokenizer
     token_path: !ref <qwen_pretrain_path>
     skip_special_tokens: True
+    version: cosyvoice3
 allowed_special: 'all'
 tokenize: !name:cosyvoice.dataset.processor.tokenize
     get_tokenizer: !ref <get_tokenizer>
@@ -146,7 +136,7 @@ filter: !name:cosyvoice.dataset.processor.filter
 resample: !name:cosyvoice.dataset.processor.resample
     resample_rate: !ref <sample_rate>
 truncate: !name:cosyvoice.dataset.processor.truncate
-    truncate_length: 24480 # must be a multiplier of hop_size
+    truncate_length: 24960 # must be a multiplier of hop_size and token_mel_ratio
 feat_extractor: !name:matcha.utils.audio.mel_spectrogram
     n_fft: 1920
     num_mels: 80
@@ -154,7 +144,7 @@ feat_extractor: !name:matcha.utils.audio.mel_spectrogram
     hop_size: 480
     win_size: 1920
     fmin: 0
-    fmax: 8000
+    fmax: null
     center: False
 compute_fbank: !name:cosyvoice.dataset.processor.compute_fbank
     feat_extractor: !ref <feat_extractor>
@@ -231,4 +221,4 @@ train_conf_gan:
     grad_clip: 5
     accum_grad: 1 # in gan training, accum_grad must be 1
     log_interval: 100
-    save_per_step: -1
+    save_per_step: -1

examples/libritts/cosyvoice3/run.sh

@@ -7,7 +7,7 @@ stop_stage=3
 
 data_url=www.openslr.org/resources/60
 data_dir=/mnt/lyuxiang.lx/data/tts/openslr/libritts
-pretrained_model_dir=../../../pretrained_models/CosyVoice3-0.5B
+pretrained_model_dir=../../../pretrained_models/Fun-CosyVoice3-0.5B
 
 if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
   echo "Data Download"
@@ -20,7 +20,8 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
   echo "Data preparation, prepare wav.scp/text/utt2spk/spk2utt"
   for x in train-clean-100 train-clean-360 train-other-500 dev-clean dev-other test-clean test-other; do
     mkdir -p data/$x
-    python local/prepare_data.py --src_dir $data_dir/LibriTTS/$x --des_dir data/$x --instruct
+    # NOTE in CosyVoice3, we add instruct in sequence
+    python local/prepare_data.py --src_dir $data_dir/LibriTTS/$x --des_dir data/$x --instruct "You are a helpful assistant.<|endofprompt|>"
   done
 fi
 
@@ -67,7 +68,6 @@ if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
   fi
   cat data/{train-clean-100,train-clean-360,train-other-500}/parquet/data.list > data/train.data.list
   cat data/{dev-clean,dev-other}/parquet/data.list > data/dev.data.list
-  # NOTE will update llm/hift training later
   for model in llm flow hifigan; do
     torchrun --nnodes=1 --nproc_per_node=$num_gpus \
         --rdzv_id=$job_id --rdzv_backend="c10d" --rdzv_endpoint="localhost:1234" \

requirements.txt

@@ -17,6 +17,7 @@ lightning==2.2.4
 matplotlib==3.7.5
 modelscope==1.20.0
 networkx==3.1
+numpy==1.26.4
 omegaconf==2.3.0
 onnx==1.16.0
 onnxruntime-gpu==1.18.0; sys_platform == 'linux'
@@ -35,6 +36,7 @@ tensorrt-cu12-libs==10.13.3.9; sys_platform == 'linux'
 torch==2.3.1
 torchaudio==2.3.1
 transformers==4.51.3
+x-transformers==2.11.24
 uvicorn==0.30.0
 wetext==0.0.4
 wget==3.2

runtime/python/fastapi/server.py

@@ -24,7 +24,7 @@ import numpy as np
 ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
 sys.path.append('{}/../../..'.format(ROOT_DIR))
 sys.path.append('{}/../../../third_party/Matcha-TTS'.format(ROOT_DIR))
-from cosyvoice.cli.cosyvoice import CosyVoice, CosyVoice2
+from cosyvoice.cli.cosyvoice import AutoModel
 from cosyvoice.utils.file_utils import load_wav
 
 app = FastAPI()
@@ -88,14 +88,8 @@ if __name__ == '__main__':
                         default=50000)
     parser.add_argument('--model_dir',
                         type=str,
-                        default='iic/CosyVoice-300M',
+                        default='iic/CosyVoice2-0.5B',
                         help='local path or modelscope repo id')
     args = parser.parse_args()
-    try:
-        cosyvoice = CosyVoice(args.model_dir)
-    except Exception:
-        try:
-            cosyvoice = CosyVoice2(args.model_dir)
-        except Exception:
-            raise TypeError('no valid model_type!')
+    cosyvoice = AutoModel(model_dir=args.model_dir)
     uvicorn.run(app, host="0.0.0.0", port=args.port)

runtime/python/grpc/server.py

@@ -25,7 +25,7 @@ import numpy as np
 ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
 sys.path.append('{}/../../..'.format(ROOT_DIR))
 sys.path.append('{}/../../../third_party/Matcha-TTS'.format(ROOT_DIR))
-from cosyvoice.cli.cosyvoice import CosyVoice, CosyVoice2
+from cosyvoice.cli.cosyvoice import AutoModel
 
 logging.basicConfig(level=logging.DEBUG,
                     format='%(asctime)s %(levelname)s %(message)s')
@@ -33,13 +33,7 @@ logging.basicConfig(level=logging.DEBUG,
 
 class CosyVoiceServiceImpl(cosyvoice_pb2_grpc.CosyVoiceServicer):
     def __init__(self, args):
-        try:
-            self.cosyvoice = CosyVoice(args.model_dir, trt_concurrent=args.max_conc)
-        except Exception:
-            try:
-                self.cosyvoice = CosyVoice2(args.model_dir, trt_concurrent=args.max_conc)
-            except Exception:
-                raise TypeError('no valid model_type!')
+        self.cosyvoice = AutoModel(model_dir=args.model_dir)
         logging.info('grpc service initialized')
 
     def Inference(self, request, context):
@@ -90,7 +84,7 @@ if __name__ == '__main__':
                         default=4)
     parser.add_argument('--model_dir',
                         type=str,
-                        default='iic/CosyVoice-300M',
+                        default='iic/CosyVoice2-0.5B',
                         help='local path or modelscope repo id')
     args = parser.parse_args()
     main()

runtime/triton_trtllm/model_repo/cosyvoice2/1/model.py

@@ -28,6 +28,7 @@ import json
 import os
 import threading
 import time
+from uuid import uuid4
 
 import numpy as np
 import torch
@@ -364,6 +365,7 @@ class TritonPythonModel:
             # Generate semantic tokens with LLM
             generated_ids_iter = self.forward_llm(input_ids)
 
+            token2wav_request_id = request_id or str(uuid4())
             if self.decoupled:
                 response_sender = request.get_response_sender()
 
@@ -392,7 +394,7 @@ class TritonPythonModel:
                         this_tts_speech_token = torch.tensor(this_tts_speech_token).unsqueeze(dim=0).to(torch.int32).to(self.device)
 
                         sub_tts_speech = self.forward_token2wav(
-                            this_tts_speech_token, request_id, prompt_speech_tokens,
+                            this_tts_speech_token, token2wav_request_id, prompt_speech_tokens,
                             prompt_speech_feat, prompt_spk_embedding, token_offset, False
                         )
 
@@ -427,7 +429,7 @@ class TritonPythonModel:
                         time.sleep(0.02)
 
                 this_tts_speech_token = torch.tensor(semantic_token_ids_arr).unsqueeze(dim=0).to(torch.int32).to(self.device)
-                sub_tts_speech = self.forward_token2wav(this_tts_speech_token, request_id, prompt_speech_tokens, prompt_speech_feat, prompt_spk_embedding, token_offset, True)
+                sub_tts_speech = self.forward_token2wav(this_tts_speech_token, token2wav_request_id, prompt_speech_tokens, prompt_speech_feat, prompt_spk_embedding, token_offset, True)
                 audio_tensor = pb_utils.Tensor.from_dlpack("waveform", to_dlpack(sub_tts_speech))
                 inference_response = pb_utils.InferenceResponse(output_tensors=[audio_tensor])
                 response_sender.send(inference_response)
@@ -441,7 +443,7 @@ class TritonPythonModel:
                 if generated_ids is None or len(generated_ids) == 0:
                     raise pb_utils.TritonModelException("Generated IDs is None or empty")
 
-                audio = self.forward_token2wav(generated_ids, request_id, prompt_speech_tokens, prompt_speech_feat, prompt_spk_embedding)
+                audio = self.forward_token2wav(generated_ids, token2wav_request_id, prompt_speech_tokens, prompt_speech_feat, prompt_spk_embedding)
 
                 # Prepare response
                 audio_tensor = pb_utils.Tensor.from_dlpack("waveform", to_dlpack(audio))

tools/make_parquet_list.py

@@ -86,7 +86,7 @@ if __name__ == "__main__":
                         help='Use Direct Preference Optimization')
     args = parser.parse_args()
 
-    utt2wav, utt2text, utt2spk = {}, {}, {}
+    utt2wav, utt2text, utt2spk, utt2instruct = {}, {}, {}, {}
     with open('{}/wav.scp'.format(args.src_dir)) as f:
         for l in f:
             l = l.replace('\n', '').split()

webui.py

@@ -57,9 +57,6 @@ def generate_audio(tts_text, mode_checkbox_group, sft_dropdown, prompt_text, pro
         prompt_wav = None
     # if instruct mode, please make sure that model is iic/CosyVoice-300M-Instruct and not cross_lingual mode
     if mode_checkbox_group in ['自然语言控制']:
-        if cosyvoice.instruct is False:
-            gr.Warning('您正在使用自然语言控制模式, {}模型不支持此模式, 请使用iic/CosyVoice-300M-Instruct模型'.format(args.model_dir))
-            yield (cosyvoice.sample_rate, default_data)
         if instruct_text == '':
             gr.Warning('您正在使用自然语言控制模式, 请输入instruct文本')
             yield (cosyvoice.sample_rate, default_data)
@@ -67,9 +64,6 @@ def generate_audio(tts_text, mode_checkbox_group, sft_dropdown, prompt_text, pro
             gr.Info('您正在使用自然语言控制模式, prompt音频/prompt文本会被忽略')
     # if cross_lingual mode, please make sure that model is iic/CosyVoice-300M and tts_text prompt_text are different language
     if mode_checkbox_group in ['跨语种复刻']:
-        if cosyvoice.instruct is True:
-            gr.Warning('您正在使用跨语种复刻模式, {}模型不支持此模式, 请使用iic/CosyVoice-300M模型'.format(args.model_dir))
-            yield (cosyvoice.sample_rate, default_data)
         if instruct_text != '':
             gr.Info('您正在使用跨语种复刻模式, instruct文本会被忽略')
         if prompt_wav is None:
@@ -167,7 +161,7 @@ if __name__ == '__main__':
                         default=8000)
     parser.add_argument('--model_dir',
                         type=str,
-                        default='pretrained_models/CosyVoice3-0.5B',
+                        default='pretrained_models/CosyVoice2-0.5B',
                         help='local path or modelscope repo id')
     args = parser.parse_args()
     cosyvoice = AutoModel(model_dir=args.model_dir)