deepgeek/CosyVoice
2
0
Fork 0
mirror of https://github.com/FunAudioLLM/CosyVoice.git synced 2026-02-05 09:59:23 +08:00
Code Issues Packages Projects Releases Wiki Activity

add llm export script

Browse Source
This commit is contained in:
lyuxiang.lx
2024-08-28 18:21:11 +08:00
parent f4e70e222c
commit 9ab298dd49
9 changed files with 34 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
cosyvoice/cli/cosyvoice.py
View File

@@ -21,7 +21,7 @@ from cosyvoice.utils.file_utils import logging
class CosyVoice:
def __init__(self, model_dir):
def __init__(self, model_dir, load_script=True):
instruct = True if '-Instruct' in model_dir else False
self.model_dir = model_dir
if not os.path.exists(model_dir):
@@ -39,6 +39,9 @@ class CosyVoice:
self.model.load('{}/llm.pt'.format(model_dir),
'{}/flow.pt'.format(model_dir),
'{}/hift.pt'.format(model_dir))
if load_script:
self.model.load_script('{}/llm.text_encoder.fp16.zip'.format(model_dir),
'{}/llm.llm.fp16.zip'.format(model_dir))
del configs
def list_avaliable_spks(self):

9
cosyvoice/cli/model.py
View File

@@ -47,11 +47,18 @@ class CosyVoiceModel:
def load(self, llm_model, flow_model, hift_model):
self.llm.load_state_dict(torch.load(llm_model, map_location=self.device))
self.llm.to(self.device).eval()
self.llm.half()
self.flow.load_state_dict(torch.load(flow_model, map_location=self.device))
self.flow.to(self.device).eval()
self.hift.load_state_dict(torch.load(hift_model, map_location=self.device))
self.hift.to(self.device).eval()
def load_script(self, llm_text_encoder_model, llm_llm_model):
llm_text_encoder = torch.jit.load(llm_text_encoder_model)
self.llm.text_encoder = llm_text_encoder
llm_llm = torch.jit.load(llm_llm_model)
self.llm.llm = llm_llm
def llm_job(self, text, text_len, prompt_text, prompt_text_len, llm_prompt_speech_token, llm_prompt_speech_token_len, llm_embedding, this_uuid):
with self.llm_context:
for i in self.llm.inference(text=text.to(self.device),
@@ -60,7 +67,7 @@ class CosyVoiceModel:
prompt_text_len=prompt_text_len.to(self.device),
prompt_speech_token=llm_prompt_speech_token.to(self.device),
prompt_speech_token_len=llm_prompt_speech_token_len.to(self.device),
embedding=llm_embedding.to(self.device),
embedding=llm_embedding.to(self.device).half(),
sampling=25,
max_token_text_ratio=30,
min_token_text_ratio=3):