logits输出有问题[Bug] about lmdeploy HOT 9 OPEN

from transformers import AutoModel, AutoTokenizer
import torch
import os
from transformers import LlamaForCausalLM, LlamaTokenizer
model_path = "NousResearch/Llama-2-7b-hf"

# 加载分词器
tokenizer = LlamaTokenizer.from_pretrained(model_path)

# 加载模型
model = LlamaForCausalLM.from_pretrained(model_path)

# 输入文本
input_text = "请根据给定的文本信息，判断是否存在的违规行为。回复:"
input_ids = tokenizer(input_text, return_tensors='pt').input_ids

# 获取logits
with torch.no_grad():
    outputs = model(input_ids)
logits1 = outputs.logits

del model
torch.cuda.empty_cache()


# turbomind
from lmdeploy import pipeline
pipe = pipeline(model_path, log_level='INFO')
g = pipe.engine.create_instance()
tok = pipe.tokenizer
inputs = '请根据给定的文本信息，判断是否存在的违规行为。回复:'
input_ids = tok.encode(inputs) 
logits2 = g.decode(input_ids).cpu()
print(logits2.shape)
# diff
rc = torch.max(torch.abs(logits1 - logits2))
print(rc)
#tensor(0.4661)

m = AutoModel.from_pretrained('/workspace/models-140/llama2/huggingface/llama-2-7b-chat', trust_remote_code=True).half().eval().cuda()
input_ids = torch.LongTensor(ids).reshape(1, -1)
with torch.inference_mode():
  out = m(input_ids.cuda())
  logits1 = out.logits.cpu()