Merge branch 'main' into bug_cpu_hang

cosyvoice/cli/model.py

@@ -58,15 +58,15 @@ class CosyVoiceModel:
         self.hift_cache_dict = {}
 
     def load(self, llm_model, flow_model, hift_model):
-        self.llm.load_state_dict(torch.load(llm_model, map_location=self.device), strict=False)
+        self.llm.load_state_dict(torch.load(llm_model, map_location=self.device), strict=True)
         self.llm.to(self.device).eval()
         if self.fp16 is True:
             self.llm.half()
-        self.flow.load_state_dict(torch.load(flow_model, map_location=self.device), strict=False)
+        self.flow.load_state_dict(torch.load(flow_model, map_location=self.device), 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()}
-        self.hift.load_state_dict(hift_state_dict, strict=False)
+        self.hift.load_state_dict(hift_state_dict, strict=True)
         self.hift.to(self.device).eval()
 
     def load_jit(self, llm_text_encoder_model, llm_llm_model, flow_encoder_model):
@@ -260,3 +260,167 @@ class CosyVoiceModel:
             self.llm_end_dict.pop(this_uuid)
             self.mel_overlap_dict.pop(this_uuid)
             self.hift_cache_dict.pop(this_uuid)
+
+
+class CosyVoice2Model:
+
+    def __init__(self,
+                 llm: torch.nn.Module,
+                 flow: torch.nn.Module,
+                 hift: torch.nn.Module):
+        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+        self.llm = llm
+        self.flow = flow
+        self.hift = hift
+        self.token_hop_len = 2 * self.flow.input_frame_rate
+        # here we fix flow encoder/decoder decoding_chunk_size, in the future we will send it as arguments, or use cache
+        self.flow.encoder.static_chunk_size = 2 * self.flow.input_frame_rate
+        self.flow.decoder.estimator.static_chunk_size = 2 * self.flow.input_frame_rate * self.flow.token_mel_ratio
+        # 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.stream_scale_factor = 1
+        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 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.to(self.device).eval()
+        self.flow.load_state_dict(torch.load(flow_model, map_location=self.device), strict=True)
+        self.flow.to(self.device).eval()
+        self.flow.decoder.fp16 = False
+        # 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()}
+        self.hift.load_state_dict(hift_state_dict, strict=True)
+        self.hift.to(self.device).eval()
+
+    def load_jit(self, flow_encoder_model):
+        flow_encoder = torch.jit.load(flow_encoder_model, map_location=self.device)
+        self.flow.encoder = flow_encoder
+
+    def load_onnx(self, flow_decoder_estimator_model):
+        import onnxruntime
+        option = onnxruntime.SessionOptions()
+        option.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
+        option.intra_op_num_threads = 1
+        providers = ['CUDAExecutionProvider' if torch.cuda.is_available() else 'CPUExecutionProvider']
+        del self.flow.decoder.estimator
+        self.flow.decoder.estimator = onnxruntime.InferenceSession(flow_decoder_estimator_model, sess_options=option, providers=providers)
+
+    def load_trt(self, flow_decoder_estimator_model):
+        del self.flow.decoder.estimator
+        import tensorrt as trt
+        with open(flow_decoder_estimator_model, 'rb') as f:
+            self.flow.decoder.estimator_engine = trt.Runtime(trt.Logger(trt.Logger.INFO)).deserialize_cuda_engine(f.read())
+        self.flow.decoder.estimator = self.flow.decoder.estimator_engine.create_execution_context()
+        self.flow.decoder.fp16 = True
+
+    def llm_job(self, text, prompt_text, llm_prompt_speech_token, llm_embedding, uuid):
+        with self.llm_context:
+            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)):
+                self.tts_speech_token_dict[uuid].append(i)
+        self.llm_end_dict[uuid] = True
+
+    def token2wav(self, token, prompt_token, prompt_feat, embedding, uuid, token_offset, finalize=False, speed=1.0):
+        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),
+                                         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
+
+    def tts(self, text, flow_embedding, llm_embedding=torch.zeros(0, 192),
+            prompt_text=torch.zeros(1, 0, dtype=torch.int32),
+            llm_prompt_speech_token=torch.zeros(1, 0, dtype=torch.int32),
+            flow_prompt_speech_token=torch.zeros(1, 0, dtype=torch.int32),
+            prompt_speech_feat=torch.zeros(1, 0, 80), stream=False, speed=1.0, **kwargs):
+        # this_uuid is used to track variables related to this inference thread
+        this_uuid = str(uuid.uuid1())
+        with self.lock:
+            self.tts_speech_token_dict[this_uuid], self.llm_end_dict[this_uuid] = [], False
+            self.hift_cache_dict[this_uuid] = None
+        p = threading.Thread(target=self.llm_job, args=(text, prompt_text, llm_prompt_speech_token, llm_embedding, this_uuid))
+        p.start()
+        if stream is True:
+            token_offset = 0
+            while True:
+                time.sleep(0.1)
+                if len(self.tts_speech_token_dict[this_uuid]) - token_offset >= self.token_hop_len + self.flow.pre_lookahead_len:
+                    this_tts_speech_token = torch.tensor(self.tts_speech_token_dict[this_uuid][:token_offset + self.token_hop_len + self.flow.pre_lookahead_len]).unsqueeze(dim=0)
+                    this_tts_speech = self.token2wav(token=this_tts_speech_token,
+                                                     prompt_token=flow_prompt_speech_token,
+                                                     prompt_feat=prompt_speech_feat,
+                                                     embedding=flow_embedding,
+                                                     uuid=this_uuid,
+                                                     token_offset=token_offset,
+                                                     finalize=False)
+                    token_offset += self.token_hop_len
+                    yield {'tts_speech': this_tts_speech.cpu()}
+                if self.llm_end_dict[this_uuid] is True and len(self.tts_speech_token_dict[this_uuid]) - token_offset < self.token_hop_len + self.flow.pre_lookahead_len:
+                    break
+            p.join()
+            # deal with remain tokens, make sure inference remain token len equals token_hop_len when cache_speech is not None
+            this_tts_speech_token = torch.tensor(self.tts_speech_token_dict[this_uuid]).unsqueeze(dim=0)
+            this_tts_speech = self.token2wav(token=this_tts_speech_token,
+                                             prompt_token=flow_prompt_speech_token,
+                                             prompt_feat=prompt_speech_feat,
+                                             embedding=flow_embedding,
+                                             uuid=this_uuid,
+                                             token_offset=token_offset,
+                                             finalize=True)
+            yield {'tts_speech': this_tts_speech.cpu()}
+        else:
+            # deal with all tokens
+            p.join()
+            this_tts_speech_token = torch.tensor(self.tts_speech_token_dict[this_uuid]).unsqueeze(dim=0)
+            this_tts_speech = self.token2wav(token=this_tts_speech_token,
+                                             prompt_token=flow_prompt_speech_token,
+                                             prompt_feat=prompt_speech_feat,
+                                             embedding=flow_embedding,
+                                             uuid=this_uuid,
+                                             token_offset=0,
+                                             finalize=True,
+                                             speed=speed)
+            yield {'tts_speech': this_tts_speech.cpu()}
+        with self.lock:
+            self.tts_speech_token_dict.pop(this_uuid)
+            self.llm_end_dict.pop(this_uuid)