he attention mask is not set and cannot be inferred from input because pad token is same as eos toke

使用llama3 推理时出现以下警告：

The attention mask is not set and cannot be inferred from input because pad token is same as eos token. As a consequence, you may observe unexpected behavior. Please pass your input’s attention_mask to obtain reliable results.

这是我的原代码

from transformers import AutoModelForCausalLM, AutoTokenizer
import transformers
import torch 
model_name = "/home/base_models/llama3-70b-instruct"

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained(model_name, use_safetensors=True)
tokenizer = AutoTokenizer.from_pretrained(model_name)
prompt="Given 17 categories: 'math, health, physics, business, biology, chemistry, computer science, economics, engineering, philosophy, other, history, geography, politics, psychology, culture, law', determine the following instruction pair belongs to which category: \
  just output the final category you choose with nothing else {\
    \"messages\": [{\"role\": \"user\", \"content\": \"Choose the best answer. Which of the following is a type of data visualization? Input: A) A graph\\nB) A heatmap\\nC) A soundtrack\"}, {\"role\": \"assistant\", \"content\": \"B) A heatmap\"}\
  }"
input_ids = tokenizer.encode(prompt, return_tensors='pt') #tokenizer.encode 仅返回input_ids
generated_text = model.generate(
    input_ids=input_ids,
    max_new_tokens=512,  # 生成文本的最大长度
    num_return_sequences=1,  # 生成多少个序列
    no_repeat_ngram_size=2,  # 避免重复的n-gram
    repetition_penalty=1.5,  # 重复惩罚系数
    top_p=0.95,  # 顶端概率值
    temperature=0.7,  # 控制随机性
    do_sample=True  # 是否使用采样；False 表示使用贪婪解码
)
generated_ids = [
        output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)]
# 解码生成的文本
output_text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
print(output_text)

重点在 input_ids = tokenizer.encode(prompt, return_tensors=‘pt’) 这一行，使用tokenizer.encode只会返回input_ids，使用tokenizer(prompt, return_tensors=‘pt’) 则会同时返回input_ids和attention_mask。

所以改为

from transformers import AutoModelForCausalLM, AutoTokenizer
import transformers
import torch 
model_name = "/home/base_models/llama3-70b-instruct"

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained(model_name, use_safetensors=True)
tokenizer = AutoTokenizer.from_pretrained(model_name)
prompt="Given 17 categories: 'math, health, physics, business, biology, chemistry, computer science, economics, engineering, philosophy, other, history, geography, politics, psychology, culture, law', determine the following instruction pair belongs to which category: \
  just output the final category you choose with nothing else {\
    \"messages\": [{\"role\": \"user\", \"content\": \"Choose the best answer. Which of the following is a type of data visualization? Input: A) A graph\\nB) A heatmap\\nC) A soundtrack\"}, {\"role\": \"assistant\", \"content\": \"B) A heatmap\"}\
  }"
inputs = tokenizer(prompt, return_tensors='pt') #tokenizer.encode 仅返回input_ids
generated_text = model.generate(
    **inputs,
    max_new_tokens=512,  # 生成文本的最大长度
    num_return_sequences=1,  # 生成多少个序列
    no_repeat_ngram_size=2,  # 避免重复的n-gram
    repetition_penalty=1.5,  # 重复惩罚系数
    top_p=0.95,  # 顶端概率值
    temperature=0.7,  # 控制随机性
    do_sample=True  # 是否使用采样；False 表示使用贪婪解码
)
generated_ids = [
        output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)]
# 解码生成的文本
output_text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
print(output_text)

即可