update

cosyvoice/cli/cosyvoice.py

@@ -19,7 +19,7 @@ from hyperpyyaml import load_hyperpyyaml
 from modelscope import snapshot_download
 import torch
 from cosyvoice.cli.frontend import CosyVoiceFrontEnd
-from cosyvoice.cli.model import CosyVoiceModel, CosyVoice2Model
+from cosyvoice.cli.model import CosyVoiceModel, CosyVoice2Model, CosyVoice3Model
 from cosyvoice.utils.file_utils import logging
 from cosyvoice.utils.class_utils import get_model_type
 
@@ -192,3 +192,43 @@ class CosyVoice2(CosyVoice):
                 logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
                 yield model_output
                 start_time = time.time()
+
+
+class CosyVoice3(CosyVoice):
+
+    def __init__(self, model_dir, load_jit=False, load_trt=False, load_vllm=False, fp16=False, trt_concurrent=1):
+        self.instruct = True if '-Instruct' in model_dir else False
+        self.model_dir = model_dir
+        self.fp16 = fp16
+        if not os.path.exists(model_dir):
+            model_dir = snapshot_download(model_dir)
+        hyper_yaml_path = '{}/cosyvoice3.yaml'.format(model_dir)
+        if not os.path.exists(hyper_yaml_path):
+            raise ValueError('{} not found!'.format(hyper_yaml_path))
+        with open(hyper_yaml_path, 'r') as f:
+            configs = load_hyperpyyaml(f, overrides={'qwen_pretrain_path': os.path.join(model_dir, 'CosyVoice-BlankEN')})
+        assert get_model_type(configs) == CosyVoice2Model, 'do not use {} for CosyVoice2 initialization!'.format(model_dir)
+        self.frontend = CosyVoiceFrontEnd(configs['get_tokenizer'],
+                                          configs['feat_extractor'],
+                                          '{}/campplus.onnx'.format(model_dir),
+                                          '{}/speech_tokenizer_v3.onnx'.format(model_dir),
+                                          '{}/spk2info.pt'.format(model_dir),
+                                          configs['allowed_special'])
+        self.sample_rate = configs['sample_rate']
+        if torch.cuda.is_available() is False and (load_jit is True or load_trt is True or fp16 is True):
+            load_jit, load_trt, fp16 = False, False, False
+            logging.warning('no cuda device, set load_jit/load_trt/fp16 to False')
+        self.model = CosyVoice3Model(configs['llm'], configs['flow'], configs['hift'], fp16)
+        self.model.load('{}/llm.pt'.format(model_dir),
+                        '{}/flow.pt'.format(model_dir),
+                        '{}/bigvgan.pt'.format(model_dir))
+        if load_vllm:
+            self.model.load_vllm('{}/vllm'.format(model_dir))
+        if load_jit:
+            self.model.load_jit('{}/flow.encoder.{}.zip'.format(model_dir, 'fp16' if self.fp16 is True else 'fp32'))
+        if load_trt:
+            self.model.load_trt('{}/flow.decoder.estimator.{}.mygpu.plan'.format(model_dir, 'fp16' if self.fp16 is True else 'fp32'),
+                                '{}/flow.decoder.estimator.fp32.onnx'.format(model_dir),
+                                trt_concurrent,
+                                self.fp16)
+        del configs

cosyvoice/cli/model.py

@@ -384,3 +384,70 @@ class CosyVoice2Model(CosyVoiceModel):
         if torch.cuda.is_available():
             torch.cuda.empty_cache()
             torch.cuda.current_stream().synchronize()
+
+
+class CosyVoice3Model(CosyVoice2Model):
+
+    def __init__(self,
+                 llm: torch.nn.Module,
+                 flow: torch.nn.Module,
+                 hift: torch.nn.Module,
+                 fp16: bool = False):
+        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+        self.llm = llm
+        self.flow = flow
+        self.hift = hift
+        self.fp16 = fp16
+        if self.fp16 is True:
+            self.llm.half()
+            self.flow.half()
+        # NOTE must matching training static_chunk_size
+        self.token_hop_len = 25
+        # hift cache
+        self.mel_cache_len = 8
+        self.source_cache_len = int(self.mel_cache_len * 480)
+        # speech fade in out
+        self.speech_window = np.hamming(2 * self.source_cache_len)
+        # rtf and decoding related
+        self.llm_context = torch.cuda.stream(torch.cuda.Stream(self.device)) if torch.cuda.is_available() else nullcontext()
+        self.lock = threading.Lock()
+        # dict used to store session related variable
+        self.tts_speech_token_dict = {}
+        self.llm_end_dict = {}
+        self.hift_cache_dict = {}
+
+    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):
+            tts_mel, _ = self.flow.inference(token=token.to(self.device),
+                                             token_len=torch.tensor([token.shape[1]], dtype=torch.int32).to(self.device),
+                                             prompt_token=prompt_token.to(self.device),
+                                             prompt_token_len=torch.tensor([prompt_token.shape[1]], dtype=torch.int32).to(self.device),
+                                             prompt_feat=prompt_feat.to(self.device),
+                                             prompt_feat_len=torch.tensor([prompt_feat.shape[1]], dtype=torch.int32).to(self.device),
+                                             embedding=embedding.to(self.device),
+                                             streaming=stream,
+                                             finalize=finalize)
+        tts_mel = tts_mel[:, :, token_offset * self.flow.token_mel_ratio:]
+        # append hift cache
+        if self.hift_cache_dict[uuid] is not None:
+            hift_cache_mel, hift_cache_source = self.hift_cache_dict[uuid]['mel'], self.hift_cache_dict[uuid]['source']
+            tts_mel = torch.concat([hift_cache_mel, tts_mel], dim=2)
+        else:
+            hift_cache_source = torch.zeros(1, 1, 0)
+        # keep overlap mel and hift cache
+        if finalize is False:
+            tts_speech, tts_source = self.hift.inference(speech_feat=tts_mel, cache_source=hift_cache_source)
+            if self.hift_cache_dict[uuid] is not None:
+                tts_speech = fade_in_out(tts_speech, self.hift_cache_dict[uuid]['speech'], self.speech_window)
+            self.hift_cache_dict[uuid] = {'mel': tts_mel[:, :, -self.mel_cache_len:],
+                                          'source': tts_source[:, :, -self.source_cache_len:],
+                                          'speech': tts_speech[:, -self.source_cache_len:]}
+            tts_speech = tts_speech[:, :-self.source_cache_len]
+        else:
+            if speed != 1.0:
+                assert self.hift_cache_dict[uuid] is None, 'speed change only support non-stream inference mode'
+                tts_mel = F.interpolate(tts_mel, size=int(tts_mel.shape[2] / speed), mode='linear')
+            tts_speech, tts_source = self.hift.inference(speech_feat=tts_mel, cache_source=hift_cache_source)
+            if self.hift_cache_dict[uuid] is not None:
+                tts_speech = fade_in_out(tts_speech, self.hift_cache_dict[uuid]['speech'], self.speech_window)
+        return tts_speech