xinference 1.1.0__py3-none-any.whl → 1.2.0__py3-none-any.whl

xinference/thirdparty/cosyvoice/cli/model.py

@@ -15,6 +15,7 @@ import torch
 import numpy as np
 import threading
 import time
+from torch.nn import functional as F
 from contextlib import nullcontext
 import uuid
 from cosyvoice.utils.common import fade_in_out
@@ -25,100 +26,134 @@ class CosyVoiceModel:
     def __init__(self,
                  llm: torch.nn.Module,
                  flow: torch.nn.Module,
-                 hift: torch.nn.Module):
+                 hift: torch.nn.Module,
+                 fp16: bool):
         self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
         self.llm = llm
         self.flow = flow
         self.hift = hift
-        self.token_min_hop_len = 100
-        self.token_max_hop_len = 200
+        self.fp16 = fp16
+        self.token_min_hop_len = 2 * self.flow.input_frame_rate
+        self.token_max_hop_len = 4 * self.flow.input_frame_rate
         self.token_overlap_len = 20
+        # here we fix set flow.decoder.estimator.static_chunk_size = 0 for compatibability
+        self.flow.decoder.estimator.static_chunk_size = 0
         # mel fade in out
-        self.mel_overlap_len = 34
+        self.mel_overlap_len = int(self.token_overlap_len / self.flow.input_frame_rate * 22050 / 256)
         self.mel_window = np.hamming(2 * self.mel_overlap_len)
         # hift cache
         self.mel_cache_len = 20
         self.source_cache_len = int(self.mel_cache_len * 256)
+        # speech fade in out
+        self.speech_window = np.hamming(2 * self.source_cache_len)
         # rtf and decoding related
         self.stream_scale_factor = 1
         assert self.stream_scale_factor >= 1, 'stream_scale_factor should be greater than 1, change it according to your actual rtf'
         self.llm_context = torch.cuda.stream(torch.cuda.Stream(self.device)) if torch.cuda.is_available() else nullcontext()
-        self.flow_hift_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.mel_overlap_dict = {}
+        self.flow_cache_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))
+        self.llm.load_state_dict(torch.load(llm_model, map_location=self.device), strict=True)
         self.llm.to(self.device).eval()
-        self.llm.half()
-        self.flow.load_state_dict(torch.load(flow_model, map_location=self.device))
+        if self.fp16 is True:
+            self.llm.half()
+        self.flow.load_state_dict(torch.load(flow_model, map_location=self.device), strict=True)
         self.flow.to(self.device).eval()
-        self.hift.load_state_dict(torch.load(hift_model, map_location=self.device))
+        # 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, llm_text_encoder_model, llm_llm_model):
-        llm_text_encoder = torch.jit.load(llm_text_encoder_model)
+    def load_jit(self, llm_text_encoder_model, llm_llm_model, flow_encoder_model):
+        assert self.fp16 is True, "we only provide fp16 jit model, set fp16=True if you want to use jit model"
+        llm_text_encoder = torch.jit.load(llm_text_encoder_model, map_location=self.device)
         self.llm.text_encoder = llm_text_encoder
-        llm_llm = torch.jit.load(llm_llm_model)
+        llm_llm = torch.jit.load(llm_llm_model, map_location=self.device)
         self.llm.llm = llm_llm
+        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 llm_job(self, text, prompt_text, llm_prompt_speech_token, llm_embedding, uuid):
+        if self.fp16 is True:
+            llm_embedding = llm_embedding.half()
         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).half(),
-                                                sampling=25,
-                                                max_token_text_ratio=30,
-                                                min_token_text_ratio=3):
+                                        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, finalize=False):
-        with self.flow_hift_context:
-            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))
-            # mel overlap fade in out
-            # if self.mel_overlap_dict[uuid] is not None:
-            #     tts_mel = fade_in_out(tts_mel, self.mel_overlap_dict[uuid], self.mel_window)
-            # append hift cache
+    def token2wav(self, token, prompt_token, prompt_feat, embedding, uuid, finalize=False, speed=1.0):
+        tts_mel, flow_cache = 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),
+                                                  flow_cache=self.flow_cache_dict[uuid])
+        self.flow_cache_dict[uuid] = flow_cache
+
+        # mel overlap fade in out
+        if self.mel_overlap_dict[uuid].shape[2] != 0:
+            tts_mel = fade_in_out(tts_mel, self.mel_overlap_dict[uuid], self.mel_window)
+        # 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:
+            self.mel_overlap_dict[uuid] = tts_mel[:, :, -self.mel_overlap_len:]
+            tts_mel = tts_mel[:, :, :-self.mel_overlap_len]
+            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:
-                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:
-                self.mel_overlap_dict[uuid] = tts_mel[:, :, -self.mel_overlap_len:]
-                tts_mel = tts_mel[:, :, :-self.mel_overlap_len]
-                tts_speech, tts_source = self.hift.inference(mel=tts_mel, cache_source=hift_cache_source)
-                self.hift_cache_dict[uuid] = {'source': tts_source[:, :, -self.source_cache_len:], 'mel': tts_mel[:, :, -self.mel_cache_len:]}
-                tts_speech = tts_speech[:, :-self.source_cache_len]
-            else:
-                tts_speech, tts_source = self.hift.inference(mel=tts_mel, cache_source=hift_cache_source)
+                tts_speech = fade_in_out(tts_speech, self.hift_cache_dict[uuid]['speech'], self.speech_window)
         return tts_speech
 
-    def inference(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, **kwargs):
+    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], self.mel_overlap_dict[this_uuid], self.hift_cache_dict[this_uuid] = [], False, None, None
+            self.tts_speech_token_dict[this_uuid], self.llm_end_dict[this_uuid] = [], False
+            self.hift_cache_dict[this_uuid] = None
+            self.mel_overlap_dict[this_uuid] = torch.zeros(1, 80, 0)
+            self.flow_cache_dict[this_uuid] = torch.zeros(1, 80, 0, 2)
         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:
@@ -126,15 +161,15 @@ class CosyVoiceModel:
             while True:
                 time.sleep(0.1)
                 if len(self.tts_speech_token_dict[this_uuid]) >= token_hop_len + self.token_overlap_len:
-                    this_tts_speech_token = torch.concat(self.tts_speech_token_dict[this_uuid][:token_hop_len + self.token_overlap_len], dim=1)
-                    with self.flow_hift_context:
-                        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,
-                                                    finalize=False)
-                    yield  {'tts_speech': this_tts_speech.cpu()}
+                    this_tts_speech_token = torch.tensor(self.tts_speech_token_dict[this_uuid][:token_hop_len + self.token_overlap_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,
+                                                     finalize=False)
+                    yield {'tts_speech': this_tts_speech.cpu()}
                     with self.lock:
                         self.tts_speech_token_dict[this_uuid] = self.tts_speech_token_dict[this_uuid][token_hop_len:]
                     # increase token_hop_len for better speech quality
@@ -143,31 +178,246 @@ class CosyVoiceModel:
                     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.concat(self.tts_speech_token_dict[this_uuid], dim=1)
-            with self.flow_hift_context:
-                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,
-                                            finalize=True)
+            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,
+                                             finalize=True)
             yield {'tts_speech': this_tts_speech.cpu()}
         else:
             # deal with all tokens
             p.join()
-            this_tts_speech_token = torch.concat(self.tts_speech_token_dict[this_uuid], dim=1)
-            with self.flow_hift_context:
-                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,
-                                            finalize=True)
+            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,
+                                             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)
             self.mel_overlap_dict.pop(this_uuid)
             self.hift_cache_dict.pop(this_uuid)
-        if torch.cuda.is_initialized():
-            torch.cuda.synchronize()
+            self.flow_cache_dict.pop(this_uuid)
+
+    def vc(self, source_speech_token, flow_prompt_speech_token, prompt_speech_feat, flow_embedding, 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] = source_speech_token.flatten().tolist(), True
+            self.hift_cache_dict[this_uuid] = None
+            self.mel_overlap_dict[this_uuid] = torch.zeros(1, 80, 0)
+            self.flow_cache_dict[this_uuid] = torch.zeros(1, 80, 0, 2)
+        if stream is True:
+            token_hop_len = self.token_min_hop_len
+            while True:
+                if len(self.tts_speech_token_dict[this_uuid]) >= token_hop_len + self.token_overlap_len:
+                    this_tts_speech_token = torch.tensor(self.tts_speech_token_dict[this_uuid][:token_hop_len + self.token_overlap_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,
+                                                     finalize=False)
+                    yield {'tts_speech': this_tts_speech.cpu()}
+                    with self.lock:
+                        self.tts_speech_token_dict[this_uuid] = self.tts_speech_token_dict[this_uuid][token_hop_len:]
+                    # increase token_hop_len for better speech quality
+                    token_hop_len = min(self.token_max_hop_len, int(token_hop_len * self.stream_scale_factor))
+                if self.llm_end_dict[this_uuid] is True and len(self.tts_speech_token_dict[this_uuid]) < token_hop_len + self.token_overlap_len:
+                    break
+            # 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,
+                                             finalize=True)
+            yield {'tts_speech': this_tts_speech.cpu()}
+        else:
+            # deal with all tokens
+            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,
+                                             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)
+            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)

xinference/thirdparty/cosyvoice/dataset/dataset.py

@@ -126,6 +126,7 @@ class DataList(IterableDataset):
 def Dataset(data_list_file,
             data_pipeline,
             mode='train',
+            gan=False,
             shuffle=True,
             partition=True,
             tts_file='',
@@ -148,13 +149,16 @@ def Dataset(data_list_file,
             tts_data = json.load(f)
         utt2lists = read_json_lists(prompt_utt2data)
         # filter unnecessary file in inference mode
-        lists = list(set([utt2lists[utt] for utt in tts_data.keys() if utt2lists[utt] in lists]))
+        lists = list({utt2lists[utt] for utt in tts_data.keys() if utt2lists[utt] in lists})
     dataset = DataList(lists,
                        shuffle=shuffle,
                        partition=partition)
     if mode == 'inference':
-        # map partial arg tts_data in inference mode
+        # map partial arg to parquet_opener func in inference mode
         data_pipeline[0] = partial(data_pipeline[0], tts_data=tts_data)
+    if gan is True:
+        # map partial arg to padding func in gan mode
+        data_pipeline[-1] = partial(data_pipeline[-1], gan=gan)
     for func in data_pipeline:
         dataset = Processor(dataset, func, mode=mode)
     return dataset