update

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