deepgeek/CosyVoice
2
0
Fork 0
mirror of https://github.com/FunAudioLLM/CosyVoice.git synced 2026-02-04 17:39:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix lint

Browse Source
This commit is contained in:
lyuxiang.lx
2025-02-06 16:07:13 +08:00
parent 24f796a2b1
commit 2a3e033ee1
17 changed files with 187 additions and 135 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
cosyvoice/bin/average_model.py
View File

@@ -75,7 +75,7 @@ def main():
print('Processing {}'.format(path))
states = torch.load(path, map_location=torch.device('cpu'))
for k in states.keys():
if k not in avg.keys():
if k not in avg.keys() and k not in ['step', 'epoch']:
avg[k] = states[k].clone()
else:
avg[k] += states[k]

1
cosyvoice/bin/export_jit.py
View File

@@ -98,5 +98,6 @@ def main():
script.save('{}/flow.encoder.fp16.zip'.format(args.model_dir))
logging.info('successfully export flow_encoder')
if __name__ == '__main__':
main()

36
cosyvoice/bin/export_onnx.py
View File

@@ -99,7 +99,7 @@ def main():
option.intra_op_num_threads = 1
providers = ['CUDAExecutionProvider' if torch.cuda.is_available() else 'CPUExecutionProvider']
estimator_onnx = onnxruntime.InferenceSession('{}/flow.decoder.estimator.fp32.onnx'.format(args.model_dir),
sess_options=option, providers=providers)
sess_options=option, providers=providers)
for _ in tqdm(range(10)):
x, mask, mu, t, spks, cond = get_dummy_input(batch_size, random.randint(16, 512), out_channels, device)
@@ -131,31 +131,33 @@ def main():
torch.onnx.export(
estimator,
(x, mask, mu, t, spks, cond,
cache['down_blocks_conv_cache'],
cache['down_blocks_kv_cache'],
cache['mid_blocks_conv_cache'],
cache['mid_blocks_kv_cache'],
cache['up_blocks_conv_cache'],
cache['up_blocks_kv_cache'],
cache['final_blocks_conv_cache']),
cache['down_blocks_conv_cache'],
cache['down_blocks_kv_cache'],
cache['mid_blocks_conv_cache'],
cache['mid_blocks_kv_cache'],
cache['up_blocks_conv_cache'],
cache['up_blocks_kv_cache'],
cache['final_blocks_conv_cache']),
'{}/flow.decoder.estimator.fp32.onnx'.format(args.model_dir),
export_params=True,
opset_version=18,
do_constant_folding=True,
input_names=['x', 'mask', 'mu', 't', 'spks', 'cond', 'down_blocks_conv_cache', 'down_blocks_kv_cache', 'mid_blocks_conv_cache', 'mid_blocks_kv_cache', 'up_blocks_conv_cache', 'up_blocks_kv_cache', 'final_blocks_conv_cache'],
output_names=['estimator_out', 'down_blocks_conv_cache_out', 'down_blocks_kv_cache_out', 'mid_blocks_conv_cache_out', 'mid_blocks_kv_cache_out', 'up_blocks_conv_cache_out', 'up_blocks_kv_cache_out', 'final_blocks_conv_cache_out'],
input_names=['x', 'mask', 'mu', 't', 'spks', 'cond', 'down_blocks_conv_cache', 'down_blocks_kv_cache', 'mid_blocks_conv_cache', 'mid_blocks_kv_cache',
'up_blocks_conv_cache', 'up_blocks_kv_cache', 'final_blocks_conv_cache'],
output_names=['estimator_out', 'down_blocks_conv_cache_out', 'down_blocks_kv_cache_out', 'mid_blocks_conv_cache_out', 'mid_blocks_kv_cache_out',
'up_blocks_conv_cache_out', 'up_blocks_kv_cache_out', 'final_blocks_conv_cache_out'],
dynamic_axes={
'x': {2: 'seq_len'},
'mask': {2: 'seq_len'},
'mu': {2: 'seq_len'},
'cond': {2: 'seq_len'},
'down_blocks_kv_cache': {3: 'seq_len'},
'mid_blocks_kv_cache': {3: 'seq_len'},
'up_blocks_kv_cache': {3: 'seq_len'},
'down_blocks_kv_cache': {3: 'cache_in_len'},
'mid_blocks_kv_cache': {3: 'cache_in_len'},
'up_blocks_kv_cache': {3: 'cache_in_len'},
'estimator_out': {2: 'seq_len'},
'down_blocks_kv_cache_out': {3: 'seq_len'},
'mid_blocks_kv_cache_out': {3: 'seq_len'},
'up_blocks_kv_cache_out': {3: 'seq_len'},
'down_blocks_kv_cache_out': {3: 'cache_out_len'},
'mid_blocks_kv_cache_out': {3: 'cache_out_len'},
'up_blocks_kv_cache_out': {3: 'cache_out_len'},
}
)
@@ -165,7 +167,7 @@ def main():
option.intra_op_num_threads = 1
providers = ['CUDAExecutionProvider' if torch.cuda.is_available() else 'CPUExecutionProvider']
estimator_onnx = onnxruntime.InferenceSession('{}/flow.decoder.estimator.fp32.onnx'.format(args.model_dir),
sess_options=option, providers=providers)
sess_options=option, providers=providers)
for _ in tqdm(range(10)):
x, mask, mu, t, spks, cond = get_dummy_input(batch_size, random.randint(16, 512), out_channels, device)

12
cosyvoice/bin/export_trt.sh
View File

@@ -7,19 +7,19 @@ MODEL_DIR=<YOUR_MODEL_DIR>
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$TRT_DIR/lib:/usr/local/cuda/lib64
# cosyvoice export
$TRT_DIR/bin/trtexec --onnx=$MODEL_DIR/flow.decoder.estimator.fp32.onnx --saveEngine=$MODEL_DIR/flow.decoder.estimator.fp32.mygpu.plan --minShapes=x:2x80x4,mask:2x1x4,mu:2x80x4,cond:2x80x4 --optShapes=x:2x80x193,mask:2x1x193,mu:2x80x193,cond:2x80x193 --maxShapes=x:2x80x6800,mask:2x1x6800,mu:2x80x6800,cond:2x80x6800 --inputIOFormats=fp32:chw,fp32:chw,fp32:chw,fp32:chw,fp32:chw,fp32:chw --outputIOFormats=fp32:chw
$TRT_DIR/bin/trtexec --onnx=$MODEL_DIR/flow.decoder.estimator.fp32.onnx --saveEngine=$MODEL_DIR/flow.decoder.estimator.fp16.mygpu.plan --fp16 --minShapes=x:2x80x4,mask:2x1x4,mu:2x80x4,cond:2x80x4 --optShapes=x:2x80x193,mask:2x1x193,mu:2x80x193,cond:2x80x193 --maxShapes=x:2x80x6800,mask:2x1x6800,mu:2x80x6800,cond:2x80x6800 --inputIOFormats=fp16:chw,fp16:chw,fp16:chw,fp16:chw,fp16:chw,fp16:chw --outputIOFormats=fp16:chw
$TRT_DIR/bin/trtexec --onnx=$MODEL_DIR/flow.decoder.estimator.fp32.onnx --saveEngine=$MODEL_DIR/flow.decoder.estimator.fp32.mygpu.plan --minShapes=x:2x80x4,mask:2x1x4,mu:2x80x4,cond:2x80x4 --optShapes=x:2x80x200,mask:2x1x200,mu:2x80x200,cond:2x80x200 --maxShapes=x:2x80x3000,mask:2x1x3000,mu:2x80x3000,cond:2x80x3000 --inputIOFormats=fp32:chw,fp32:chw,fp32:chw,fp32:chw,fp32:chw,fp32:chw --outputIOFormats=fp32:chw
$TRT_DIR/bin/trtexec --onnx=$MODEL_DIR/flow.decoder.estimator.fp32.onnx --saveEngine=$MODEL_DIR/flow.decoder.estimator.fp16.mygpu.plan --fp16 --minShapes=x:2x80x4,mask:2x1x4,mu:2x80x4,cond:2x80x4 --optShapes=x:2x80x200,mask:2x1x200,mu:2x80x200,cond:2x80x200 --maxShapes=x:2x80x3000,mask:2x1x3000,mu:2x80x3000,cond:2x80x3000 --inputIOFormats=fp16:chw,fp16:chw,fp16:chw,fp16:chw,fp16:chw,fp16:chw --outputIOFormats=fp16:chw
# cosyvoice2 export with cache
$TRT_DIR/bin/trtexec --onnx=$MODEL_DIR/flow.decoder.estimator.fp32.onnx --saveEngine=$MODEL_DIR/flow.decoder.estimator.fp32.mygpu.plan \
--minShapes=x:2x80x4,mask:2x1x4,mu:2x80x4,cond:2x80x4,down_blocks_kv_cache:1x4x2x0x512x2,mid_blocks_kv_cache:12x4x2x0x512x2,up_blocks_kv_cache:1x4x2x0x512x2 \
--optShapes=x:2x80x193,mask:2x1x193,mu:2x80x193,cond:2x80x193,down_blocks_kv_cache:1x4x2x193x512x2,mid_blocks_kv_cache:12x4x2x193x512x2,up_blocks_kv_cache:1x4x2x193x512x2 \
--maxShapes=x:2x80x6800,mask:2x1x6800,mu:2x80x6800,cond:2x80x6800,down_blocks_kv_cache:1x4x2x200x512x2,mid_blocks_kv_cache:12x4x2x200x512x2,up_blocks_kv_cache:1x4x2x200x512x2 \
--optShapes=x:2x80x200,mask:2x1x200,mu:2x80x200,cond:2x80x200,down_blocks_kv_cache:1x4x2x100x512x2,mid_blocks_kv_cache:12x4x2x100x512x2,up_blocks_kv_cache:1x4x2x100x512x2 \
--maxShapes=x:2x80x1500,mask:2x1x1500,mu:2x80x1500,cond:2x80x1500,down_blocks_kv_cache:1x4x2x200x512x2,mid_blocks_kv_cache:12x4x2x200x512x2,up_blocks_kv_cache:1x4x2x200x512x2 \
--inputIOFormats=fp32:chw,fp32:chw,fp32:chw,fp32:chw,fp32:chw,fp32:chw,fp32:chw,fp32:chw,fp32:chw,fp32:chw,fp32:chw,fp32:chw,fp32:chw \
--outputIOFormats=fp32:chw,fp32:chw,fp32:chw,fp32:chw,fp32:chw,fp32:chw,fp32:chw,fp32:chw
$TRT_DIR/bin/trtexec --onnx=$MODEL_DIR/flow.decoder.estimator.fp32.onnx --saveEngine=$MODEL_DIR/flow.decoder.estimator.fp16.mygpu.plan --fp16 \
--minShapes=x:2x80x4,mask:2x1x4,mu:2x80x4,cond:2x80x4,down_blocks_kv_cache:1x4x2x0x512x2,mid_blocks_kv_cache:12x4x2x0x512x2,up_blocks_kv_cache:1x4x2x0x512x2 \
--optShapes=x:2x80x193,mask:2x1x193,mu:2x80x193,cond:2x80x193,down_blocks_kv_cache:1x4x2x193x512x2,mid_blocks_kv_cache:12x4x2x193x512x2,up_blocks_kv_cache:1x4x2x193x512x2 \
--maxShapes=x:2x80x6800,mask:2x1x6800,mu:2x80x6800,cond:2x80x6800,down_blocks_kv_cache:1x4x2x200x512x2,mid_blocks_kv_cache:12x4x2x200x512x2,up_blocks_kv_cache:1x4x2x200x512x2 \
--optShapes=x:2x80x200,mask:2x1x200,mu:2x80x200,cond:2x80x200,down_blocks_kv_cache:1x4x2x100x512x2,mid_blocks_kv_cache:12x4x2x100x512x2,up_blocks_kv_cache:1x4x2x100x512x2 \
--maxShapes=x:2x80x1500,mask:2x1x1500,mu:2x80x1500,cond:2x80x1500,down_blocks_kv_cache:1x4x2x200x512x2,mid_blocks_kv_cache:12x4x2x200x512x2,up_blocks_kv_cache:1x4x2x200x512x2 \
--inputIOFormats=fp16:chw,fp16:chw,fp16:chw,fp16:chw,fp16:chw,fp16:chw,fp16:chw,fp16:chw,fp16:chw,fp16:chw,fp16:chw,fp16:chw,fp16:chw \
--outputIOFormats=fp16:chw,fp16:chw,fp16:chw,fp16:chw,fp16:chw,fp16:chw,fp16:chw,fp16:chw

3
cosyvoice/bin/inference.py
View File

@@ -78,6 +78,7 @@ def main():
tts_file=args.tts_text, prompt_utt2data=args.prompt_utt2data)
test_data_loader = DataLoader(test_dataset, batch_size=None, num_workers=0)
sample_rate = configs['sample_rate']
del configs
os.makedirs(args.result_dir, exist_ok=True)
fn = os.path.join(args.result_dir, 'wav.scp')
@@ -113,7 +114,7 @@ def main():
tts_speeches = torch.concat(tts_speeches, dim=1)
tts_key = '{}_{}'.format(utts[0], tts_index[0])
tts_fn = os.path.join(args.result_dir, '{}.wav'.format(tts_key))
torchaudio.save(tts_fn, tts_speeches, sample_rate=configs['sample_rate'], backend='soundfile')
torchaudio.save(tts_fn, tts_speeches, sample_rate=sample_rate, backend='soundfile')
f.write('{} {}\n'.format(tts_key, tts_fn))
f.flush()
f.close()