in route.py [0:0]
def __init__(self, model= ''):
self.model = model
model = model.lower().replace('_','').replace('-','')
if 'qwen2' in model:
self.tag ='qwen2'
elif 'llama3' in model:
self.tag ='llama3'
elif 'llama2' in model:
self.tag ='llam2'
elif 'deepseek' in model:
self.tag ='deepseek'
elif 'mistral' in model:
self.tag ='mistral'
elif 'codellama' in model:
self.tag = 'codellama'
else:
raise TypeError(f"Unexpect model: {model}.")
self.llm = LLM(model=self.model,
seed=123,
gpu_memory_utilization=0.9,
tensor_parallel_size=args.gpus,
trust_remote_code=True,
)
self.tokenizer = self.llm.get_tokenizer()