CuteGPT is an open-source conversational language model that supports both Chinese and English, developed by [Fudan University KnowledgeWorks Laboratory](http://kw.fudan.edu.cn/). It is based on the original Llama model structure, and has a scale of 13B (13 billion) parameters. It can perform int8 precision inference on a single 3090 graphics card. CuteGPT expands the Chinese vocabulary and performs pre-training on the Llama model, improving its ability to understand Chinese. Subsequently, it is fine-tuned with conversational instructions to enhance the model's ability to understand instructions.
Based on the KW-CuteGPT-7b version, KW-CuteGPT-13b has improved accuracy in knowledge, understanding of complex instructions, ability to comprehend long texts, reasoning ability, faithful question answering, and other capabilities. Currently, the KW-CuteGPT-13b version model outperforms the majority of models of similar scale in certain evaluation tasks.

```python
from transformers import LlamaForCausalLM, LlamaTokenizer
from peft import PeftModel
import torch
```

* The prompt template for inference

```python
overall_instruction = "你是复旦大学知识工场实验室训练出来的语言模型CuteGPT。给定任务描述，请给出对应请求的回答。\n"
def generate_prompt(query, history, input=None):
    prompt = overall_instruction
    for i, (old_query, response) in enumerate(history):
        prompt += "Q: {}\nA: {}\n".format(old_query, response)
    prompt += "Q: {}\nA: ".format(query)
    return prompt
```

* Load model, tokenizer, here we use lora version of checkpoint
  * w/o 8bit quantization
    ```python
    model_name = "XuYipei/kw-cutegpt-13b-base"
    LORA_WEIGHTS = "Abbey4799/kw-cutegpt-13b-ift-lora"
    tokenizer = LlamaTokenizer.from_pretrained(LORA_WEIGHTS)
    model = LlamaForCausalLM.from_pretrained(
        model_name,
        torch_dtype=torch.float16,
        device_map="auto",
    )
    model.eval()
    model = PeftModel.from_pretrained(model, LORA_WEIGHTS).to(torch.float16)
    device = torch.device("cuda")
    ```
  * w/ 8bit quantization (The performance of the model will experience some degradation after quantization.)
    ```python
    model_name = "XuYipei/kw-cutegpt-13b-base"
    LORA_WEIGHTS = "Abbey4799/kw-cutegpt-13b-ift-lora"
    tokenizer = LlamaTokenizer.from_pretrained(LORA_WEIGHTS)
    model = LlamaForCausalLM.from_pretrained(
        model_name,
        load_in_8bit=True,
        torch_dtype=torch.float16,
        device_map="auto",
    )
    model.eval()
    model = PeftModel.from_pretrained(model, LORA_WEIGHTS)
    device = torch.device("cuda")
    ```

* Inference

```python
history = []
queries = ['请推荐五本名著，依次列出作品名、作者','再来三本呢？']
memory_limit = 3 # the number of (query, response) to remember
for query in queries:
    prompt = generate_prompt(query, history)
    print(prompt)

    input_ids = tokenizer(prompt, return_tensors="pt", padding=False, truncation=False, add_special_tokens=False)
    input_ids = input_ids["input_ids"].to(device)

    with torch.no_grad():
        outputs=model.generate(
                input_ids=input_ids,
                top_p=0.8,
                top_k=50,
                repetition_penalty=1.1,
                max_new_tokens = 256,
                early_stopping = True,
                eos_token_id = tokenizer.convert_tokens_to_ids('<s>'),
                pad_token_id = tokenizer.eos_token_id,
                min_length = input_ids.shape[1] + 1
        )
    s = outputs[0][input_ids.shape[1]:]
    response=tokenizer.decode(s)
    response = response.replace('<s>', '').replace('<end>', '').replace('</s>', '')
    print(response)
    history.append((query, response))
    history = history[-memory_limit:]
```