---
license: apache-2.0
base_model:
- internlm/internlm2_5-7b-chat
---
# Converted Qwen2 from InternLM2.5-7B-Chat

## Descritpion
This is a converted model from [InternLM2.5-7B-Chat](https://huggingface.co/internlm/internlm2_5-7b-chat) to __Qwen2__ format. This conversion allows you to use InternLM2.5-7B-Chat as if it were a Qwen2 model, which is convenient for some *inference use cases*. The __precision__ is __excatly the same__ as the original model. 

## Usage
You can load the model using the `Qwen2ForCausalLM` class as shown below:
```python
device = "cpu" # cpu is exacatly the same
attn_impl = 'eager' # the attention implementation to use
meta_instruction = ("You are an AI assistant whose name is InternLM (书生·浦语).\n"
"- InternLM (书生·浦语) is a conversational language model that is developed by Shanghai AI Laboratory "
"(上海人工智能实验室). It is designed to be helpful, honest, and harmless.\n"
"- InternLM (书生·浦语) can understand and communicate fluently in the language chosen by the user such "
"as English and 中文."
)
prompt1 = "介绍下你自己"
prompt2 = "介绍下上海人工智能实验室"

def build_inputs(tokenizer, query: str, history: List[Tuple[str, str]] = None, meta_instruction=meta_instruction):
    if history is None:
        history = []
    if tokenizer.add_bos_token:
        prompt = ""
    else:
        prompt = tokenizer.bos_token
    if meta_instruction:
        prompt += f"""<|im_start|>system\n{meta_instruction}<|im_end|>\n"""
    for record in history:
        prompt += f"""<|im_start|>user\n{record[0]}<|im_end|>\n<|im_start|>assistant\n{record[1]}<|im_end|>\n"""
    prompt += f"""<|im_start|>user\n{query}<|im_end|>\n<|im_start|>assistant\n"""
    return tokenizer([prompt], return_tensors="pt")

@torch.inference_mode()
def chat(
    model: Union[AutoModelForCausalLM, Qwen2ForCausalLM],
    tokenizer,
    query: str,
    history: Optional[List[Tuple[str, str]]] = None,
    streamer: Optional[BaseStreamer] = None,
    max_new_tokens: int = 1024,
    do_sample: bool = True,
    temperature: float = 0.8,
    top_p: float = 0.8,
    meta_instruction: str = meta_instruction,
    **kwargs,
):
    if history is None:
        history = []
    inputs = build_inputs(tokenizer, query, history, meta_instruction)
    inputs = {k: v.to(model.device) for k, v in inputs.items() if torch.is_tensor(v)}
    # also add end-of-assistant token in eos token id to avoid unnecessary generation
    eos_token_id = [tokenizer.eos_token_id, tokenizer.convert_tokens_to_ids(["<|im_end|>"])[0]]
    outputs = model.generate(
        **inputs,
        streamer=streamer,
        max_new_tokens=max_new_tokens,
        do_sample=do_sample,
        temperature=temperature,
        top_p=top_p,
        eos_token_id=eos_token_id,
        **kwargs,
    )
    outputs = outputs[0].cpu().tolist()[len(inputs["input_ids"][0]) :]
    response = tokenizer.decode(outputs, skip_special_tokens=True)
    response = response.split("<|im_end|>")[0]
    history = history + [(query, response)]
    return response, history
    

# use the official tokenizer
tokenizer = AutoTokenizer.from_pretrained("silence09/InternLM2.5-7B-Chat-Converted-Qwen2", trust_remote_code=True)
# use the converted LlaMA model
qwen2_model = Qwen2ForCausalLM.from_pretrained(
    "silence09/InternLM2.5-7B-Chat-Converted-Qwen2",
    torch_dtype='auto',
    attn_implementation=attn_impl).to(device)
qwen2_model.eval()
response_qwen2_and_splitfunc_1, history = chat(qwen2_model, tokenizer, prompt1, history=[], do_sample=False)
print(f"User Input: {prompt1}\nConverted LlaMA Response: {response_qwen2_and_splitfunc_1}")

response_qwen2_and_splitfunc_2, history = chat(qwen2_model, tokenizer, prompt2, history=history, do_sample=False)
print(f"User Input: {prompt2}\nConverted LlaMA Response: {response_qwen2_and_splitfunc_2}")

```

## Precision Guarantee
To comare result with the original model, you can use this [code](https://github.com/silencelamb/naked_llama/blob/main/hf_example/hf_internlm_7b_qwen2_compare.py)

## More Info
It was converted using the python script available at [this repository](https://github.com/silencelamb/naked_llama/blob/main/hf_example/convert_internlm_to_qwen_hf.py)