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

update func export_codec_vllm

Browse Source
This commit is contained in:
雾聪
2025-02-26 16:48:21 +08:00
parent f6a18ee07a
commit f280558bcb
3 changed files with 5 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
cosyvoice/cli/cosyvoice.py
View File

@@ -156,7 +156,7 @@ class CosyVoice2(CosyVoice):
self.model.export_codec_vllm(''.join([model_dir, '/codec_vllm_model']))
engine_args = EngineArgs(model=''.join([model_dir, '/codec_vllm_model']),
skip_tokenizer_init=True,
gpu_memory_utilization=0.1)
gpu_memory_utilization=0.2)
self.vllm_codec_engine = LLMEngine.from_engine_args(engine_args)
self.model.vllm_codec_engine = self.vllm_codec_engine

3
cosyvoice/cli/model.py
View File

@@ -347,6 +347,9 @@ class CosyVoice2Model(CosyVoiceModel):
self.llm.llm.model.to(dtype)
tmp_vocab_size = self.llm.llm.model.config.vocab_size
tmp_tie_embedding = self.llm.llm.model.config.tie_word_embeddings
del self.llm.llm.model.generation_config.eos_token_id
del self.llm.llm.model.config.bos_token_id
del self.llm.llm.model.config.eos_token_id
self.llm.llm.model.config.vocab_size = pad_vocab_size
self.llm.llm.model.config.tie_word_embeddings = False
self.llm.llm.model.config.use_bias = True

2
cosyvoice/llm/llm.py
View File

@@ -343,7 +343,7 @@ class Qwen2LM(TransformerLM):
max_tokens=max_len)
request_id = uuid.uuid4()
vllm_codec_engine.add_request(request_id,
{"prompt_embeds": lm_input.to(torch.bfloat16).to(device)},
{"prompt_embeds": lm_input.squeeze(0).to(torch.bfloat16).to(device)},
sampling_params)
## generator
out_token_ids = []