deepgeek/CosyVoice
2
0
Fork 0
mirror of https://github.com/FunAudioLLM/CosyVoice.git synced 2026-02-05 18:09:24 +08:00
Code Issues Packages Projects Releases Wiki Activity

初步合并vllm支持，异步推理的通道处理还存在bug

Browse Source
This commit is contained in:
qihua
2025-03-07 20:26:19 +08:00
parent fd45708e4b
commit 90b666ea20
5 changed files with 658 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

43
cosyvoice/cli/cosyvoice.py
View File

@@ -19,7 +19,7 @@ from hyperpyyaml import load_hyperpyyaml
from modelscope import snapshot_download
import torch
from cosyvoice.cli.frontend import CosyVoiceFrontEnd
from cosyvoice.cli.model import CosyVoiceModel, CosyVoice2Model
from cosyvoice.cli.model import CosyVoiceModel, CosyVoice2Model, VllmCosyVoice2Model
from cosyvoice.utils.file_utils import logging
from cosyvoice.utils.class_utils import get_model_type
@@ -63,6 +63,9 @@ class CosyVoice:
spks = list(self.frontend.spk2info.keys())
return spks
def add_spk_info(self, spk_id, spk_info):
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)
@@ -88,6 +91,22 @@ class CosyVoice:
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)
@@ -126,7 +145,7 @@ class CosyVoice:
class CosyVoice2(CosyVoice):
def __init__(self, model_dir, load_jit=False, load_trt=False, fp16=False):
def __init__(self, model_dir, load_jit=False, load_trt=False, fp16=False, use_vllm=False):
self.instruct = True if '-Instruct' in model_dir else False
self.model_dir = model_dir
self.fp16 = fp16
@@ -145,7 +164,14 @@ class CosyVoice2(CosyVoice):
if torch.cuda.is_available() is False and (load_jit is True or load_trt is True or fp16 is True):
load_jit, load_trt, fp16 = False, False, False
logging.warning('no cuda device, set load_jit/load_trt/fp16 to False')
self.model = CosyVoice2Model(configs['llm'], configs['flow'], configs['hift'], fp16)
if use_vllm:
try:
self.model = VllmCosyVoice2Model(model_dir, configs['flow'], configs['hift'], fp16)
except Exception as e:
logging.warning(f'use vllm inference failed. \n{e}')
raise e
else:
self.model = CosyVoice2Model(configs['llm'], configs['flow'], configs['hift'], fp16)
self.model.load('{}/llm.pt'.format(model_dir),
'{}/flow.pt'.format(model_dir),
'{}/hift.pt'.format(model_dir))
@@ -171,3 +197,14 @@ class CosyVoice2(CosyVoice):
logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
yield model_output
start_time = time.time()
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
start_time = time.time()