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

Revert "mv AsyncLLMEngine init to CosyVoice2"

Browse Source 
This reverts commit 9b3f351496.
This commit is contained in:
雾聪
2025-03-21 16:17:35 +08:00
parent 9b3f351496
commit 96950745a6
2 changed files with 21 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
cosyvoice/cli/cosyvoice.py
View File

@@ -166,29 +166,7 @@ class CosyVoice2(CosyVoice):
logging.warning('no cuda device, set load_jit/load_trt/fp16 to False') logging.warning('no cuda device, set load_jit/load_trt/fp16 to False')
if use_vllm: if use_vllm:
try: try:
os.environ["VLLM_USE_V1"] = '1'
from vllm import AsyncLLMEngine
from vllm.engine.arg_utils import AsyncEngineArgs
# EngineArgs
ENGINE_ARGS = {
"block_size": 16,
"swap_space": 0,
# "enforce_eager": True,
"gpu_memory_utilization": 0.4,
"max_num_batched_tokens": 1024,
"max_model_len": 1024,
"max_num_seqs": 256,
"disable_log_requests": True,
"disable_log_stats": True,
"dtype": "bfloat16"
}
self.model = VllmCosyVoice2Model(model_dir, configs['flow'], configs['hift'], fp16) self.model = VllmCosyVoice2Model(model_dir, configs['flow'], configs['hift'], fp16)
engine_args = AsyncEngineArgs(
model=model_dir,
**ENGINE_ARGS,
)
self.llm_engine: AsyncLLMEngine = AsyncLLMEngine.from_engine_args(engine_args)
self.model.llm_engine = self.llm_engine
except Exception as e: except Exception as e:
logging.warning(f'use vllm inference failed. \n{e}') logging.warning(f'use vllm inference failed. \n{e}')
raise e raise e

22
cosyvoice/llm/llm_vllm.py
View File

@@ -31,6 +31,20 @@ from vllm.sampling_params import SamplingParams
from cosyvoice.llm.vllm_use_cosyvoice2_model import CosyVoice2Model as CosyVoice2LLM from cosyvoice.llm.vllm_use_cosyvoice2_model import CosyVoice2Model as CosyVoice2LLM
ModelRegistry.register_model("CosyVoice2Model", CosyVoice2LLM) ModelRegistry.register_model("CosyVoice2Model", CosyVoice2LLM)
# EngineArgs
ENGINE_ARGS = {
"block_size": 16,
"swap_space": 0,
# "enforce_eager": True,
"gpu_memory_utilization": 0.4,
"max_num_batched_tokens": 1024,
"max_model_len": 1024,
"max_num_seqs": 256,
"disable_log_requests": True,
"disable_log_stats": True,
"dtype": "float16"
}
from vllm.sampling_params import RequestOutputKind from vllm.sampling_params import RequestOutputKind
# SamplingParams # SamplingParams
SAMPLING_PARAMS = { SAMPLING_PARAMS = {
@@ -58,7 +72,13 @@ class VllmQwen2LM(Qwen2LM):
self.fp16 = False self.fp16 = False
self.half = lambda: None self.half = lambda: None
self.mix_ratio = mix_ratio self.mix_ratio = mix_ratio
self.llm_engine = None # ---------------------------------------------
# vllm engine 的参数配置
engine_args = AsyncEngineArgs(
model=model_dir,
**ENGINE_ARGS,
)
self.llm_engine: AsyncLLMEngine = AsyncLLMEngine.from_engine_args(engine_args)
self.speech_token_size = 6564 # 6561 + 3 self.speech_token_size = 6564 # 6561 + 3
self.llm_token_size = 151936 # llm vocab_size self.llm_token_size = 151936 # llm vocab_size