deepgeek/CosyVoice
2
0
Fork 0
mirror of https://github.com/FunAudioLLM/CosyVoice.git synced 2026-02-05 01:49:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

use automodel

Browse Source
This commit is contained in:
lyuxiang.lx
2025-12-09 15:15:05 +00:00
parent 56d9876037
commit 0c65d3c7ab
8 changed files with 56 additions and 88 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
cosyvoice/utils/file_utils.py
View File

@@ -92,29 +92,14 @@ def convert_onnx_to_trt(trt_model, trt_kwargs, onnx_model, fp16):
def export_cosyvoice2_vllm(model, model_path, device):
if os.path.exists(model_path):
return
pad_to = DEFAULT_VOCAB_PADDING_SIZE = 64
vocab_size = model.speech_embedding.num_embeddings
feature_size = model.speech_embedding.embedding_dim
pad_vocab_size = ((vocab_size + pad_to - 1) // pad_to) * pad_to
dtype = torch.bfloat16
# lm_head
use_bias = True if model.llm_decoder.bias is not None else False
new_lm_head = torch.nn.Linear(in_features=feature_size, out_features=pad_vocab_size, bias=use_bias)
with torch.no_grad():
new_lm_head.weight[:vocab_size] = model.llm_decoder.weight
new_lm_head.weight[vocab_size:] = 0
if use_bias is True:
new_lm_head.bias[:vocab_size] = model.llm_decoder.bias
new_lm_head.bias[vocab_size:] = 0
model.llm.model.lm_head = new_lm_head
new_codec_embed = torch.nn.Linear(in_features=feature_size, out_features=pad_vocab_size)
model.llm.model.lm_head = model.llm_decoder
# embed_tokens
embed_tokens = model.llm.model.model.embed_tokens
with torch.no_grad():
new_codec_embed.weight[:vocab_size] = model.speech_embedding.weight
new_codec_embed.weight[vocab_size:] = 0
model.llm.model.set_input_embeddings(new_codec_embed)
model.llm.model.set_input_embeddings(model.speech_embedding)
model.llm.model.to(device)
model.llm.model.to(dtype)
tmp_vocab_size = model.llm.model.config.vocab_size
@@ -122,14 +107,12 @@ def export_cosyvoice2_vllm(model, model_path, device):
del model.llm.model.generation_config.eos_token_id
del model.llm.model.config.bos_token_id
del model.llm.model.config.eos_token_id
model.llm.model.config.vocab_size = pad_vocab_size
model.llm.model.config.vocab_size = model.speech_embedding.num_embeddings
model.llm.model.config.tie_word_embeddings = False
model.llm.model.config.use_bias = use_bias
model.llm.model.save_pretrained(model_path)
if use_bias is True:
os.system('sed -i s@Qwen2ForCausalLM@CosyVoice2ForCausalLM@g {}/config.json'.format(os.path.abspath(model_path)))
else:
os.system('sed -i s@Qwen2ForCausalLM@Qwen2ForCausalLM@g {}/config.json'.format(os.path.abspath(model_path)))
model.llm.model.config.vocab_size = tmp_vocab_size
model.llm.model.config.tie_word_embeddings = tmp_tie_embedding
model.llm.model.set_input_embeddings(embed_tokens)