mirror of
https://github.com/FunAudioLLM/CosyVoice.git
synced 2026-02-05 18:09:24 +08:00
Enhance CosyVoice with CUDA stream management and estimator handling
- Introduced a queue-based system for managing CUDA streams to improve inference performance. - Updated inference methods to utilize CUDA streams for asynchronous processing. - Added an EstimatorWrapper class to manage TensorRT estimators, allowing for efficient execution context handling. - Modified model loading functions to support estimator count configuration. - Improved logging and performance tracking during inference operations.
This commit is contained in:
禾息
@@ -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:
|
||||
|
||||
@@ -54,11 +54,18 @@ class CosyVoice:
|
||||
'{}/llm.llm.{}.zip'.format(model_dir, 'fp16' if self.fp16 is True else 'fp32'),
|
||||
'{}/flow.encoder.{}.zip'.format(model_dir, 'fp16' if self.fp16 is True else 'fp32'))
|
||||
if load_trt:
|
||||
self.estimator_count = configs['flow']['decoder']['estimator'].get('estimator_count', 1)
|
||||
self.model.load_trt('{}/flow.decoder.estimator.{}.mygpu.plan'.format(model_dir, 'fp16' if self.fp16 is True else 'fp32'),
|
||||
'{}/flow.decoder.estimator.fp32.onnx'.format(model_dir),
|
||||
self.fp16)
|
||||
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())
|
||||
return spks
|
||||
@@ -67,80 +74,104 @@ 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):
|
||||
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
|
||||
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):
|
||||
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
|
||||
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"""
|
||||
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 = 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):
|
||||
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
|
||||
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):
|
||||
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)
|
||||
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)
|
||||
start_time = time.time()
|
||||
logging.info('synthesis text {}'.format(i))
|
||||
for model_output in self.model.tts(**model_input, stream=stream, speed=speed):
|
||||
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()
|
||||
|
||||
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()
|
||||
cuda_stream.synchronize()
|
||||
self.stream_pool.put(cuda_stream)
|
||||
|
||||
|
||||
class CosyVoice2(CosyVoice):
|
||||
@@ -178,33 +209,48 @@ class CosyVoice2(CosyVoice):
|
||||
if load_jit:
|
||||
self.model.load_jit('{}/flow.encoder.{}.zip'.format(model_dir, 'fp16' if self.fp16 is True else 'fp32'))
|
||||
if load_trt:
|
||||
self.estimator_count = configs['flow']['decoder']['estimator'].get('estimator_count', 1)
|
||||
self.model.load_trt('{}/flow.decoder.estimator.{}.mygpu.plan'.format(model_dir, 'fp16' if self.fp16 is True else 'fp32'),
|
||||
'{}/flow.decoder.estimator.fp32.onnx'.format(model_dir),
|
||||
self.fp16)
|
||||
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):
|
||||
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
|
||||
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)
|
||||
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):
|
||||
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
|
||||
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)
|
||||
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)
|
||||
|
||||
Reference in New Issue
Block a user