xinference 1.5.0.post2__py3-none-any.whl → 1.6.0__py3-none-any.whl

xinference/thirdparty/cosyvoice/cli/model.py

@@ -11,6 +11,8 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import os
+from typing import Generator
 import torch
 import numpy as np
 import threading
@@ -19,6 +21,7 @@ from torch.nn import functional as F
 from contextlib import nullcontext
 import uuid
 from cosyvoice.utils.common import fade_in_out
+from cosyvoice.utils.file_utils import convert_onnx_to_trt
 
 
 class CosyVoiceModel:
@@ -27,17 +30,18 @@ class CosyVoiceModel:
                  llm: torch.nn.Module,
                  flow: torch.nn.Module,
                  hift: torch.nn.Module,
-                 fp16: bool):
+                 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()
         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 = int(self.token_overlap_len / self.flow.input_frame_rate * 22050 / 256)
         self.mel_window = np.hamming(2 * self.mel_overlap_len)
@@ -61,8 +65,6 @@ class CosyVoiceModel:
     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()
-        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()
         # in case hift_model is a hifigan model
@@ -71,7 +73,6 @@ class CosyVoiceModel:
         self.hift.to(self.device).eval()
 
     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, map_location=self.device)
@@ -79,39 +80,62 @@ class CosyVoiceModel:
         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']
+    def load_trt(self, flow_decoder_estimator_model, flow_decoder_onnx_model, fp16):
+        assert torch.cuda.is_available(), 'tensorrt only supports gpu!'
+        if not os.path.exists(flow_decoder_estimator_model):
+            convert_onnx_to_trt(flow_decoder_estimator_model, self.get_trt_kwargs(), flow_decoder_onnx_model, fp16)
+        if os.path.getsize(flow_decoder_estimator_model) == 0:
+            raise ValueError('{} is empty file, delete it and export again!'.format(flow_decoder_estimator_model))
         del self.flow.decoder.estimator
-        self.flow.decoder.estimator = onnxruntime.InferenceSession(flow_decoder_estimator_model, sess_options=option, providers=providers)
+        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())
+        assert self.flow.decoder.estimator_engine is not None, 'failed to load trt {}'.format(flow_decoder_estimator_model)
+        self.flow.decoder.estimator = self.flow.decoder.estimator_engine.create_execution_context()
+
+    def get_trt_kwargs(self):
+        min_shape = [(2, 80, 4), (2, 1, 4), (2, 80, 4), (2, 80, 4)]
+        opt_shape = [(2, 80, 200), (2, 1, 200), (2, 80, 200), (2, 80, 200)]
+        max_shape = [(2, 80, 3000), (2, 1, 3000), (2, 80, 3000), (2, 80, 3000)]
+        input_names = ["x", "mask", "mu", "cond"]
+        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):
-        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)):
-                self.tts_speech_token_dict[uuid].append(i)
+        with self.llm_context, torch.cuda.amp.autocast(self.fp16):
+            if isinstance(text, Generator):
+                assert isinstance(self, CosyVoice2Model), 'streaming input text is only implemented for CosyVoice2!'
+                for i in 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)):
+                    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)):
+                    self.tts_speech_token_dict[uuid].append(i)
+        self.llm_end_dict[uuid] = True
+
+    def vc_job(self, source_speech_token, uuid):
+        self.tts_speech_token_dict[uuid] = source_speech_token.flatten().tolist()
         self.llm_end_dict[uuid] = True
 
     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
+        with torch.cuda.amp.autocast(self.fp16):
+            tts_mel, self.flow_cache_dict[uuid] = 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])
 
         # mel overlap fade in out
         if self.mel_overlap_dict[uuid].shape[2] != 0:
@@ -142,11 +166,11 @@ class CosyVoiceModel:
                 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),
+    def tts(self, text=torch.zeros(1, 0, dtype=torch.int32), flow_embedding=torch.zeros(0, 192), 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):
+            prompt_speech_feat=torch.zeros(1, 0, 80), source_speech_token=torch.zeros(1, 0, dtype=torch.int32), 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:
@@ -154,7 +178,10 @@ class CosyVoiceModel:
             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))
+        if source_speech_token.shape[1] == 0:
+            p = threading.Thread(target=self.llm_job, args=(text, prompt_text, llm_prompt_speech_token, llm_embedding, this_uuid))
+        else:
+            p = threading.Thread(target=self.vc_job, args=(source_speech_token, this_uuid))
         p.start()
         if stream is True:
             token_hop_len = self.token_min_hop_len
@@ -204,143 +231,89 @@ class CosyVoiceModel:
             self.mel_overlap_dict.pop(this_uuid)
             self.hift_cache_dict.pop(this_uuid)
             self.flow_cache_dict.pop(this_uuid)
+        torch.cuda.empty_cache()
 
-    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:
+class CosyVoice2Model(CosyVoiceModel):
 
     def __init__(self,
                  llm: torch.nn.Module,
                  flow: torch.nn.Module,
-                 hift: torch.nn.Module):
+                 hift: torch.nn.Module,
+                 fp16: bool = False,
+                 use_flow_cache: bool = False):
         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
+        self.fp16 = fp16
+        self.use_flow_cache = use_flow_cache
+        if self.fp16 is True:
+            self.llm.half()
+            self.flow.half()
+        # stream related params, check examples/libritts/cosyvoice2/conf/cosyvoice2.yaml
+        self.token_hop_len = 25
+        self.flow_decoder_required_cache_size = 0 if use_flow_cache is False else 1 * self.token_hop_len * 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.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), 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 init_flow_cache(self):
+        encoder_cache = {'offset': 0,
+                         'pre_lookahead_layer_conv2_cache': torch.zeros(1, 512, 2).to(self.device),
+                         'encoders_kv_cache': torch.zeros(6, 1, 8, 0, 64 * 2).to(self.device),
+                         'upsample_offset': 0,
+                         'upsample_conv_cache': torch.zeros(1, 512, 4).to(self.device),
+                         'upsample_kv_cache': torch.zeros(4, 1, 8, 0, 64 * 2).to(self.device)}
+        decoder_cache = {'offset': 0,
+                         'down_blocks_conv_cache': torch.zeros(10, 1, 2, 832, 2).to(self.device),
+                         'down_blocks_kv_cache': torch.zeros(10, 1, 4, 2, self.flow_decoder_required_cache_size, 512, 2).to(self.device),
+                         'mid_blocks_conv_cache': torch.zeros(10, 12, 2, 512, 2).to(self.device),
+                         'mid_blocks_kv_cache': torch.zeros(10, 12, 4, 2, self.flow_decoder_required_cache_size, 512, 2).to(self.device),
+                         'up_blocks_conv_cache': torch.zeros(10, 1, 2, 1024, 2).to(self.device),
+                         'up_blocks_kv_cache': torch.zeros(10, 1, 4, 2, self.flow_decoder_required_cache_size, 512, 2).to(self.device),
+                         'final_blocks_conv_cache': torch.zeros(10, 2, 256, 2).to(self.device)}
+        if self.fp16 is True:
+            for cache in [encoder_cache, decoder_cache]:
+                for k, v in cache.items():
+                    if isinstance(v, torch.Tensor):
+                        cache[k] = v.half()
+        cache = {'encoder_cache': encoder_cache, 'decoder_cache': decoder_cache}
+        return cache
 
     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 get_trt_kwargs(self):
+        min_shape = [(2, 80, 4), (2, 1, 4), (2, 80, 4), (2, 80, 4), (1, 4, 2, 0, 512, 2), (12, 4, 2, 0, 512, 2), (1, 4, 2, 0, 512, 2)]
+        opt_shape = [(2, 80, 200), (2, 1, 200), (2, 80, 200), (2, 80, 200), (1, 4, 2, 100, 512, 2), (12, 4, 2, 100, 512, 2), (1, 4, 2, 100, 512, 2)]
+        max_shape = [(2, 80, 1500), (2, 1, 1500), (2, 80, 1500), (2, 80, 1500), (1, 4, 2, 200, 512, 2), (12, 4, 2, 200, 512, 2), (1, 4, 2, 200, 512, 2)]
+        input_names = ["x", "mask", "mu", "cond", 'down_blocks_kv_cache', 'mid_blocks_kv_cache', 'up_blocks_kv_cache']
+        assert self.use_flow_cache is True, "get_trt_kwargs is set for flow cache mode. If you want to use trt with use_flow_cache=False, please set higher max_shape"
+        return {'min_shape': min_shape, 'opt_shape': opt_shape, 'max_shape': max_shape, 'input_names': input_names}
 
-    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:]
+    def token2wav(self, token, prompt_token, prompt_feat, embedding, uuid, finalize=False, speed=1.0):
+        with torch.cuda.amp.autocast(self.fp16):
+            tts_mel, self.flow_cache_dict[uuid] = 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),
+                                                                      cache=self.flow_cache_dict[uuid],
+                                                                      finalize=finalize)
         # 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']
@@ -365,34 +338,44 @@ class CosyVoice2Model:
                 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),
+    def tts(self, text=torch.zeros(1, 0, dtype=torch.int32), flow_embedding=torch.zeros(0, 192), 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):
+            prompt_speech_feat=torch.zeros(1, 0, 80), source_speech_token=torch.zeros(1, 0, dtype=torch.int32), 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))
+            self.flow_cache_dict[this_uuid] = self.init_flow_cache()
+        if source_speech_token.shape[1] == 0:
+            p = threading.Thread(target=self.llm_job, args=(text, prompt_text, llm_prompt_speech_token, llm_embedding, this_uuid))
+        else:
+            p = threading.Thread(target=self.vc_job, args=(source_speech_token, this_uuid))
         p.start()
         if stream is True:
-            token_offset = 0
+            assert self.use_flow_cache is True, "set use_flow_cache=True if you want to use stream inference to avoid OOM"
+            # NOTE in cache mode, trim flow_prompt to same size as flow_decoder_required_cache_size
+            flow_prompt_speech_token = flow_prompt_speech_token[:, -int(self.flow_decoder_required_cache_size / self.flow.token_mel_ratio):]
+            prompt_speech_feat = prompt_speech_feat[:, -self.flow_decoder_required_cache_size:]
             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)
+                if len(self.tts_speech_token_dict[this_uuid]) >= self.token_hop_len + self.flow.pre_lookahead_len:
+                    this_tts_speech_token = torch.tensor(self.tts_speech_token_dict[this_uuid][: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
+                    # NOTE in cache inference mode, we only use flow_prompt_speech_token/prompt_speech_feat in first chunk
+                    flow_prompt_speech_token = torch.zeros(1, 0, dtype=torch.int32).to(self.device)
+                    prompt_speech_feat = torch.zeros(1, 0, 80).to(self.device)
                     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:
+                    with self.lock:
+                        self.tts_speech_token_dict[this_uuid] = self.tts_speech_token_dict[this_uuid][self.token_hop_len:]
+                if self.llm_end_dict[this_uuid] is True and len(self.tts_speech_token_dict[this_uuid]) < 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
@@ -402,11 +385,11 @@ class CosyVoice2Model:
                                              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
+            assert self.use_flow_cache is False, "set use_flow_cache=False for nonstream inference"
             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,
@@ -414,10 +397,12 @@ class CosyVoice2Model:
                                              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)
+            self.hift_cache_dict.pop(this_uuid)
+            self.flow_cache_dict.pop(this_uuid)
+        torch.cuda.empty_cache()

xinference/thirdparty/cosyvoice/dataset/processor.py

@@ -20,8 +20,8 @@ import torch
 import torchaudio
 from torch.nn.utils.rnn import pad_sequence
 import torch.nn.functional as F
+import pyworld as pw
 
-torchaudio.set_audio_backend('soundfile')
 
 AUDIO_FORMAT_SETS = {'flac', 'mp3', 'm4a', 'ogg', 'opus', 'wav', 'wma'}
 
@@ -179,7 +179,7 @@ def compute_fbank(data,
         yield sample
 
 
-def compute_f0(data, pitch_extractor, mode='train'):
+def compute_f0(data, sample_rate, hop_size, mode='train'):
     """ Extract f0
 
         Args:
@@ -188,15 +188,19 @@ def compute_f0(data, pitch_extractor, mode='train'):
         Returns:
             Iterable[{key, feat, label}]
     """
+    frame_period = hop_size * 1000 / sample_rate
     for sample in data:
         assert 'sample_rate' in sample
         assert 'speech' in sample
         assert 'utt' in sample
         assert 'text_token' in sample
         waveform = sample['speech']
-        mat = pitch_extractor(waveform).transpose(1, 2)
-        mat = F.interpolate(mat, size=sample['speech_feat'].shape[0], mode='linear')
-        sample['pitch_feat'] = mat[0, 0]
+        _f0, t = pw.harvest(waveform.squeeze(dim=0).numpy().astype('double'), sample_rate, frame_period=frame_period)
+        if sum(_f0 != 0) < 5:  # this happens when the algorithm fails
+            _f0, t = pw.dio(waveform.squeeze(dim=0).numpy().astype('double'), sample_rate, frame_period=frame_period)  # if harvest fails, try dio
+        f0 = pw.stonemask(waveform.squeeze(dim=0).numpy().astype('double'), _f0, t, sample_rate)
+        f0 = F.interpolate(torch.from_numpy(f0).view(1, 1, -1), size=sample['speech_feat'].shape[0], mode='linear').view(-1)
+        sample['pitch_feat'] = f0
         yield sample