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

Merge pull request #360 from FunAudioLLM/dev/lyuxiang.lx

Browse Source 
set onnx to false as last chunk rtf unstable
This commit is contained in:
Xiang Lyu
2024-09-06 17:12:23 +08:00
committed by GitHub
parent bcda6d807c 122df8c420
commit c901a12789
3 changed files with 45 additions and 51 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
.github/workflows/lint.yml vendored
View File

@@ -2,6 +2,7 @@ name: Lint
on: on:
pull_request: pull_request:
push:
jobs: jobs:
quick-checks: quick-checks:

2
cosyvoice/cli/cosyvoice.py
View File

@@ -23,7 +23,7 @@ from cosyvoice.utils.file_utils import logging
class CosyVoice: class CosyVoice:
def __init__(self, model_dir, load_jit=True, load_onnx=True): def __init__(self, model_dir, load_jit=True, load_onnx=False):
instruct = True if '-Instruct' in model_dir else False instruct = True if '-Instruct' in model_dir else False
self.model_dir = model_dir self.model_dir = model_dir
if not os.path.exists(model_dir): if not os.path.exists(model_dir):

7
cosyvoice/cli/model.py
View File

@@ -43,7 +43,6 @@ class CosyVoiceModel:
self.stream_scale_factor = 1 self.stream_scale_factor = 1
assert self.stream_scale_factor >= 1, 'stream_scale_factor should be greater than 1, change it according to your actual rtf' assert self.stream_scale_factor >= 1, 'stream_scale_factor should be greater than 1, change it according to your actual rtf'
self.llm_context = torch.cuda.stream(torch.cuda.Stream(self.device)) if torch.cuda.is_available() else nullcontext() self.llm_context = torch.cuda.stream(torch.cuda.Stream(self.device)) if torch.cuda.is_available() else nullcontext()
self.flow_hift_context = torch.cuda.stream(torch.cuda.Stream(self.device)) if torch.cuda.is_available() else nullcontext()
self.lock = threading.Lock() self.lock = threading.Lock()
# dict used to store session related variable # dict used to store session related variable
self.tts_speech_token_dict = {} self.tts_speech_token_dict = {}
@@ -93,7 +92,6 @@ class CosyVoiceModel:
self.llm_end_dict[uuid] = True self.llm_end_dict[uuid] = True
def token2wav(self, token, prompt_token, prompt_feat, embedding, uuid, finalize=False): def token2wav(self, token, prompt_token, prompt_feat, embedding, uuid, finalize=False):
with self.flow_hift_context:
tts_mel = self.flow.inference(token=token.to(self.device), tts_mel = self.flow.inference(token=token.to(self.device),
token_len=torch.tensor([token.shape[1]], dtype=torch.int32).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=prompt_token.to(self.device),
@@ -139,7 +137,6 @@ class CosyVoiceModel:
time.sleep(0.1) time.sleep(0.1)
if len(self.tts_speech_token_dict[this_uuid]) >= token_hop_len + self.token_overlap_len: if len(self.tts_speech_token_dict[this_uuid]) >= token_hop_len + self.token_overlap_len:
this_tts_speech_token = torch.concat(self.tts_speech_token_dict[this_uuid][:token_hop_len + self.token_overlap_len], dim=1) this_tts_speech_token = torch.concat(self.tts_speech_token_dict[this_uuid][:token_hop_len + self.token_overlap_len], dim=1)
with self.flow_hift_context:
this_tts_speech = self.token2wav(token=this_tts_speech_token, this_tts_speech = self.token2wav(token=this_tts_speech_token,
prompt_token=flow_prompt_speech_token, prompt_token=flow_prompt_speech_token,
prompt_feat=prompt_speech_feat, prompt_feat=prompt_speech_feat,
@@ -156,7 +153,6 @@ class CosyVoiceModel:
p.join() p.join()
# deal with remain tokens, make sure inference remain token len equals token_hop_len when cache_speech is not None # 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.concat(self.tts_speech_token_dict[this_uuid], dim=1) this_tts_speech_token = torch.concat(self.tts_speech_token_dict[this_uuid], dim=1)
with self.flow_hift_context:
this_tts_speech = self.token2wav(token=this_tts_speech_token, this_tts_speech = self.token2wav(token=this_tts_speech_token,
prompt_token=flow_prompt_speech_token, prompt_token=flow_prompt_speech_token,
prompt_feat=prompt_speech_feat, prompt_feat=prompt_speech_feat,
@@ -168,7 +164,6 @@ class CosyVoiceModel:
# deal with all tokens # deal with all tokens
p.join() p.join()
this_tts_speech_token = torch.concat(self.tts_speech_token_dict[this_uuid], dim=1) this_tts_speech_token = torch.concat(self.tts_speech_token_dict[this_uuid], dim=1)
with self.flow_hift_context:
this_tts_speech = self.token2wav(token=this_tts_speech_token, this_tts_speech = self.token2wav(token=this_tts_speech_token,
prompt_token=flow_prompt_speech_token, prompt_token=flow_prompt_speech_token,
prompt_feat=prompt_speech_feat, prompt_feat=prompt_speech_feat,
@@ -181,5 +176,3 @@ class CosyVoiceModel:
self.llm_end_dict.pop(this_uuid) self.llm_end_dict.pop(this_uuid)
self.mel_overlap_dict.pop(this_uuid) self.mel_overlap_dict.pop(this_uuid)
self.hift_cache_dict.pop(this_uuid) self.hift_cache_dict.pop(this_uuid)
if torch.cuda.is_available():
torch.cuda.synchronize()