From 7f4c9a2c6454a1bb2a76e105069293a8370825c8 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E7=A6=BE=E6=81=AF?= Date: Wed, 16 Apr 2025 14:15:14 +0800 Subject: [PATCH] Refactor CosyVoice inference methods to streamline CUDA stream management - Removed the queue-based stream pool and integrated direct CUDA stream usage for improved performance. - Simplified inference methods by eliminating unnecessary synchronization and stream management code. - Enhanced logging for better tracking of synthesis operations and performance metrics. - Updated the model class to support CUDA stream context management, ensuring efficient resource utilization during inference. --- cosyvoice/cli/cosyvoice.py | 216 +++++++++++++------------------- cosyvoice/cli/model.py | 247 +++++++++++++++++++++---------------- 2 files changed, 226 insertions(+), 237 deletions(-) diff --git a/cosyvoice/cli/cosyvoice.py b/cosyvoice/cli/cosyvoice.py index 4c3b881..8606530 100644 --- a/cosyvoice/cli/cosyvoice.py +++ b/cosyvoice/cli/cosyvoice.py @@ -22,7 +22,7 @@ from cosyvoice.cli.frontend import CosyVoiceFrontEnd from cosyvoice.cli.model import CosyVoiceModel, CosyVoice2Model, VllmCosyVoice2Model from cosyvoice.utils.file_utils import logging from cosyvoice.utils.class_utils import get_model_type -import queue + class CosyVoice: @@ -60,11 +60,6 @@ class CosyVoice: self.fp16, self.estimator_count) del configs - thread_count = 10 - self.stream_pool = queue.Queue(maxsize=thread_count) - for _ in range(thread_count): - self.stream_pool.put(torch.cuda.Stream(self.device)) - def list_available_spks(self): spks = list(self.frontend.spk2info.keys()) @@ -74,104 +69,80 @@ class CosyVoice: self.frontend.add_spk_info(spk_id, spk_info) def inference_sft(self, tts_text, spk_id, stream=False, speed=1.0, text_frontend=True): - cuda_stream = self.stream_pool.get() - with torch.cuda.stream(cuda_stream): - for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)): - model_input = self.frontend.frontend_sft(i, spk_id) - start_time = time.time() - logging.info('synthesis text {}'.format(i)) - for model_output in self.model.tts(**model_input, stream=stream, speed=speed): - speech_len = model_output['tts_speech'].shape[1] / self.sample_rate - logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len)) - yield model_output - start_time = time.time() - cuda_stream.synchronize() - self.stream_pool.put(cuda_stream) - - def inference_zero_shot(self, tts_text, prompt_text, prompt_speech_16k, stream=False, speed=1.0, text_frontend=True): - cuda_stream = self.stream_pool.get() - with torch.cuda.stream(cuda_stream): - prompt_text = self.frontend.text_normalize(prompt_text, split=False, text_frontend=text_frontend) - for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)): - if (not isinstance(i, Generator)) and len(i) < 0.5 * len(prompt_text): - logging.warning('synthesis text {} too short than prompt text {}, this may lead to bad performance'.format(i, prompt_text)) - model_input = self.frontend.frontend_zero_shot(i, prompt_text, prompt_speech_16k, self.sample_rate) - start_time = time.time() - logging.info('synthesis text {}'.format(i)) - for model_output in self.model.tts(**model_input, stream=stream, speed=speed): - speech_len = model_output['tts_speech'].shape[1] / self.sample_rate - logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len)) - yield model_output - start_time = time.time() - cuda_stream.synchronize() - self.stream_pool.put(cuda_stream) - - def inference_zero_shot_by_spk_id(self, tts_text, spk_id, stream=False, speed=1.0, text_frontend=True): - """使用预定义的说话人执行 zero_shot 推理""" - cuda_stream = self.stream_pool.get() - with torch.cuda.stream(cuda_stream): - for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)): - model_input = self.frontend.frontend_zero_shot_by_spk_id(i, spk_id) - start_time = time.time() - last_time = start_time - chunk_index = 0 - logging.info('synthesis text {}'.format(i)) - for model_output in self.model.tts(**model_input, stream=stream, speed=speed): - speech_len = model_output['tts_speech'].shape[1] / self.sample_rate - logging.info('yield speech index:{}, len {:.2f}, rtf {:.3f}, cost {:.3f}s, all cost time {:.3f}s'.format( - chunk_index, speech_len, (time.time()-last_time)/speech_len, time.time()-last_time, time.time()-start_time)) - yield model_output - last_time = time.time() - chunk_index += 1 - cuda_stream.synchronize() - self.stream_pool.put(cuda_stream) - - def inference_cross_lingual(self, tts_text, prompt_speech_16k, stream=False, speed=1.0, text_frontend=True): - cuda_stream = self.stream_pool.get() - with torch.cuda.stream(cuda_stream): - for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)): - model_input = self.frontend.frontend_cross_lingual(i, prompt_speech_16k, self.sample_rate) - start_time = time.time() - logging.info('synthesis text {}'.format(i)) - for model_output in self.model.tts(**model_input, stream=stream, speed=speed): - speech_len = model_output['tts_speech'].shape[1] / self.sample_rate - logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len)) - yield model_output - start_time = time.time() - cuda_stream.synchronize() - self.stream_pool.put(cuda_stream) - - def inference_instruct(self, tts_text, spk_id, instruct_text, stream=False, speed=1.0, text_frontend=True): - cuda_stream = self.stream_pool.get() - with torch.cuda.stream(cuda_stream): - assert isinstance(self.model, CosyVoiceModel), 'inference_instruct is only implemented for CosyVoice!' - if self.instruct is False: - raise ValueError('{} do not support instruct inference'.format(self.model_dir)) - instruct_text = self.frontend.text_normalize(instruct_text, split=False, text_frontend=text_frontend) - for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)): - model_input = self.frontend.frontend_instruct(i, spk_id, instruct_text) - start_time = time.time() - logging.info('synthesis text {}'.format(i)) - for model_output in self.model.tts(**model_input, stream=stream, speed=speed): - speech_len = model_output['tts_speech'].shape[1] / self.sample_rate - logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len)) - yield model_output - start_time = time.time() - cuda_stream.synchronize() - self.stream_pool.put(cuda_stream) - - def inference_vc(self, source_speech_16k, prompt_speech_16k, stream=False, speed=1.0): - cuda_stream = self.stream_pool.get() - with torch.cuda.stream(cuda_stream): - model_input = self.frontend.frontend_vc(source_speech_16k, prompt_speech_16k, self.sample_rate) + for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)): + model_input = self.frontend.frontend_sft(i, spk_id) start_time = time.time() - for model_output in self.model.vc(**model_input, stream=stream, speed=speed): + logging.info('synthesis text {}'.format(i)) + for model_output in self.model.tts(**model_input, stream=stream, speed=speed): speech_len = model_output['tts_speech'].shape[1] / self.sample_rate logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len)) yield model_output start_time = time.time() - cuda_stream.synchronize() - self.stream_pool.put(cuda_stream) + + def inference_zero_shot(self, tts_text, prompt_text, prompt_speech_16k, stream=False, speed=1.0, text_frontend=True): + prompt_text = self.frontend.text_normalize(prompt_text, split=False, text_frontend=text_frontend) + for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)): + if (not isinstance(i, Generator)) and len(i) < 0.5 * len(prompt_text): + logging.warning('synthesis text {} too short than prompt text {}, this may lead to bad performance'.format(i, prompt_text)) + model_input = self.frontend.frontend_zero_shot(i, prompt_text, prompt_speech_16k, self.sample_rate) + start_time = time.time() + logging.info('synthesis text {}'.format(i)) + for model_output in self.model.tts(**model_input, stream=stream, speed=speed): + speech_len = model_output['tts_speech'].shape[1] / self.sample_rate + logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len)) + yield model_output + start_time = time.time() + + def inference_zero_shot_by_spk_id(self, tts_text, spk_id, stream=False, speed=1.0, text_frontend=True): + """使用预定义的说话人执行 zero_shot 推理""" + for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)): + model_input = self.frontend.frontend_zero_shot_by_spk_id(i, spk_id) + start_time = time.time() + last_time = start_time + chunk_index = 0 + logging.info('synthesis text {}'.format(i)) + for model_output in self.model.tts(**model_input, stream=stream, speed=speed): + speech_len = model_output['tts_speech'].shape[1] / self.sample_rate + logging.info('yield speech index:{}, len {:.2f}, rtf {:.3f}, cost {:.3f}s, all cost time {:.3f}s'.format( + chunk_index, speech_len, (time.time()-last_time)/speech_len, time.time()-last_time, time.time()-start_time)) + yield model_output + last_time = time.time() + chunk_index += 1 + + def inference_cross_lingual(self, tts_text, prompt_speech_16k, stream=False, speed=1.0, text_frontend=True): + for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)): + model_input = self.frontend.frontend_cross_lingual(i, prompt_speech_16k, self.sample_rate) + start_time = time.time() + logging.info('synthesis text {}'.format(i)) + for model_output in self.model.tts(**model_input, stream=stream, speed=speed): + speech_len = model_output['tts_speech'].shape[1] / self.sample_rate + logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len)) + yield model_output + start_time = time.time() + + def inference_instruct(self, tts_text, spk_id, instruct_text, stream=False, speed=1.0, text_frontend=True): + assert isinstance(self.model, CosyVoiceModel), 'inference_instruct is only implemented for CosyVoice!' + if self.instruct is False: + raise ValueError('{} do not support instruct inference'.format(self.model_dir)) + instruct_text = self.frontend.text_normalize(instruct_text, split=False, text_frontend=text_frontend) + for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)): + model_input = self.frontend.frontend_instruct(i, spk_id, instruct_text) + start_time = time.time() + logging.info('synthesis text {}'.format(i)) + for model_output in self.model.tts(**model_input, stream=stream, speed=speed): + speech_len = model_output['tts_speech'].shape[1] / self.sample_rate + logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len)) + yield model_output + start_time = time.time() + + def inference_vc(self, source_speech_16k, prompt_speech_16k, stream=False, speed=1.0): + model_input = self.frontend.frontend_vc(source_speech_16k, prompt_speech_16k, self.sample_rate) + start_time = time.time() + for model_output in self.model.vc(**model_input, stream=stream, speed=speed): + speech_len = model_output['tts_speech'].shape[1] / self.sample_rate + logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len)) + yield model_output + start_time = time.time() class CosyVoice2(CosyVoice): @@ -215,42 +186,29 @@ class CosyVoice2(CosyVoice): self.fp16, self.estimator_count) del configs - thread_count = 10 - self.stream_pool = queue.Queue(maxsize=thread_count) - for _ in range(thread_count): - self.stream_pool.put(torch.cuda.Stream(self.device)) def inference_instruct(self, *args, **kwargs): raise NotImplementedError('inference_instruct is not implemented for CosyVoice2!') def inference_instruct2(self, tts_text, instruct_text, prompt_speech_16k, stream=False, speed=1.0, text_frontend=True): - cuda_stream = self.stream_pool.get() - with torch.cuda.stream(cuda_stream): - assert isinstance(self.model, CosyVoice2Model), 'inference_instruct2 is only implemented for CosyVoice2!' - for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)): - model_input = self.frontend.frontend_instruct2(i, instruct_text, prompt_speech_16k, self.sample_rate) + assert isinstance(self.model, CosyVoice2Model), 'inference_instruct2 is only implemented for CosyVoice2!' + for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)): + model_input = self.frontend.frontend_instruct2(i, instruct_text, prompt_speech_16k, self.sample_rate) + start_time = time.time() + logging.info('synthesis text {}'.format(i)) + for model_output in self.model.tts(**model_input, stream=stream, speed=speed): + speech_len = model_output['tts_speech'].shape[1] / self.sample_rate + logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len)) + yield model_output start_time = time.time() - logging.info('synthesis text {}'.format(i)) - for model_output in self.model.tts(**model_input, stream=stream, speed=speed): - speech_len = model_output['tts_speech'].shape[1] / self.sample_rate - logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len)) - yield model_output - start_time = time.time() - cuda_stream.synchronize() - self.stream_pool.put(cuda_stream) def inference_instruct2_by_spk_id(self, tts_text, instruct_text, spk_id, stream=False, speed=1.0, text_frontend=True): - cuda_stream = self.stream_pool.get() - with torch.cuda.stream(cuda_stream): - assert isinstance(self.model, CosyVoice2Model), 'inference_instruct2 is only implemented for CosyVoice2!' - for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)): - model_input = self.frontend.frontend_instruct2_by_spk_id(i, instruct_text, spk_id) + for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)): + model_input = self.frontend.frontend_instruct2_by_spk_id(i, instruct_text, spk_id) + start_time = time.time() + logging.info('synthesis text {}'.format(i)) + for model_output in self.model.tts(**model_input, stream=stream, speed=speed): + speech_len = model_output['tts_speech'].shape[1] / self.sample_rate + logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len)) + yield model_output start_time = time.time() - logging.info('synthesis text {}'.format(i)) - for model_output in self.model.tts(**model_input, stream=stream, speed=speed): - speech_len = model_output['tts_speech'].shape[1] / self.sample_rate - logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len)) - yield model_output - start_time = time.time() - cuda_stream.synchronize() - self.stream_pool.put(cuda_stream) diff --git a/cosyvoice/cli/model.py b/cosyvoice/cli/model.py index 769dc92..d72816a 100644 --- a/cosyvoice/cli/model.py +++ b/cosyvoice/cli/model.py @@ -23,6 +23,7 @@ import uuid from cosyvoice.utils.common import fade_in_out from cosyvoice.utils.file_utils import convert_onnx_to_trt from cosyvoice.flow.flow_matching import EstimatorWrapper +import queue class CosyVoiceModel: @@ -66,6 +67,12 @@ class CosyVoiceModel: self.flow_cache_dict = {} self.hift_cache_dict = {} + self.stream_context_pool = queue.Queue() + for _ in range(10): + self.stream_context_pool.put(torch.cuda.stream(torch.cuda.Stream(self.device)) if torch.cuda.is_available() else nullcontext()) + + self.is_cuda_available = torch.cuda.is_available() + 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() @@ -166,63 +173,70 @@ class CosyVoiceModel: 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 - 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: - token_hop_len = self.token_min_hop_len - 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.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 - 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, - 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, - 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) - self.flow_cache_dict.pop(this_uuid) - torch.cuda.empty_cache() + + stream_context = self.stream_context_pool.get() + with stream_context: + + 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 + 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: + token_hop_len = self.token_min_hop_len + 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.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 + 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, + 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, + 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) + self.flow_cache_dict.pop(this_uuid) + + self.synchronize_stream() + self.stream_context_pool.put(stream_context) + 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 @@ -278,6 +292,10 @@ class CosyVoiceModel: self.hift_cache_dict.pop(this_uuid) torch.cuda.empty_cache() + def synchronize_stream(self): + if self.is_cuda_available: + torch.cuda.current_stream().synchronize() + class CosyVoice2Model(CosyVoiceModel): @@ -314,6 +332,12 @@ class CosyVoice2Model(CosyVoiceModel): self.llm_end_dict = {} self.hift_cache_dict = {} + self.stream_context_pool = queue.Queue() + for _ in range(10): + self.stream_context_pool.put(torch.cuda.stream(torch.cuda.Stream(self.device)) if torch.cuda.is_available() else nullcontext()) + + self.is_cuda_available = torch.cuda.is_available() + def load_jit(self, flow_encoder_model): flow_encoder = torch.jit.load(flow_encoder_model, map_location=self.device) self.flow.encoder = flow_encoder @@ -359,57 +383,64 @@ class CosyVoice2Model(CosyVoiceModel): 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) - torch.cuda.empty_cache() + self.synchronize_stream() + stream_context = self.stream_context_pool.get() + with torch.cuda.stream(stream_context): + + 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) + + self.synchronize_stream() + self.stream_context_pool.put(stream_context) + torch.cuda.empty_cache() class VllmCosyVoice2Model(CosyVoice2Model):