codefuse-ai / CodeFuse-DeepSeek-33B-4bits

Model Card for CodeFuse-DeepSeek-33B-4bits

[中文] [English]

Model Description

CodeFuse-DeepSeek-33B-4bits is the 4-bit quantized version of CodeFuse-DeepSeek-33B which is a 33B Code-LLM finetuned by QLoRA on multiple code-related tasks on the base model DeepSeek-Coder-33B.

After undergoing 4-bit quantization, the CodeFuse-DeepSeek-33B-4bits model can be loaded on either a single A10 (24GB VRAM) or an RTX 4090 (24GB VRAM). Moreover, the quantized model still achieves an impressive accuracy of 78.05% on the HumanEval pass@1 metric.

News and Updates

🔥🔥🔥 2024-01-12 CodeFuse-DeepSeek-33B-4bits has been released. Despite the quantization process, the model still achieves a remarkable 78.05% accuracy (greedy decoding) on the HumanEval pass@1 metric.

🔥🔥🔥 2024-01-12 CodeFuse-DeepSeek-33B has been released, achieving a pass@1 (greedy decoding) score of 78.65% on HumanEval.

🔥🔥 2023-11-10 CodeFuse-CodeGeeX2-6B has been released, achieving a pass@1 (greedy decoding) score of 45.12% on HumanEval, which is a 9.22% increase compared to CodeGeeX2 35.9%.

🔥🔥 2023-10-20 CodeFuse-QWen-14B technical documentation has been released. For those interested, please refer to the CodeFuse article on our WeChat official account via the provided link.( https://mp.weixin.qq.com/s/PCQPkvbvfxSPzsqjOILCDw )

🔥🔥 2023-10-16 CodeFuse-QWen-14B has been released, achieving a pass@1 (greedy decoding) score of 48.78% on HumanEval, which is a 16% increase compared to Qwen-14b's 32.3%.

🔥🔥 2023-09-27 CodeFuse-StarCoder-15B has been released, achieving a pass@1 (greedy decoding) score of 54.9% on HumanEval, which is a 21% increase compared to StarCoder's 33.6%.

🔥🔥🔥 2023-09-26 We are pleased to announce the release of the 4-bit quantized version of CodeFuse-CodeLlama-34B . Despite the quantization process, the model still achieves a remarkable 73.8% accuracy (greedy decoding) on the HumanEval pass@1 metric.

🔥🔥🔥 2023-09-11 CodeFuse-CodeLlama34B has achieved 74.4% of pass@1 (greedy decoding) on HumanEval, which is SOTA results for open-sourced LLMs at present.

Code Community

Homepage : 🏡 https://github.com/codefuse-ai ( Please give us your support with a Star🌟 + Fork🚀 + Watch👀 )

• If you wish to fine-tune the model yourself, you can visit ✨ MFTCoder ✨✨

• If you wish to deploy the model yourself, you can visit ✨ FasterTransformer4CodeFuse ✨✨

• If you wish to see a demo of the model, you can visit ✨ CodeFuse Demo ✨✨

Performance

Requirements

• python>=3.8
• pytorch>=2.0.0
• transformers>=4.33.2
• Sentencepiece
• auto_gptq
• CUDA 11.4
  

Inference String Format

The inference string is a concatenated string formed by combining conversation data (system, human and bot contents) in the training data format. It is used as input during the inference process. Here are examples of prompts used to request the model:

Multi-Round with System Prompt:

"""
<s>system
System instruction
<s>human
Human 1st round input
<s>bot
Bot 1st round output<｜end▁of▁sentence｜>
<s>human
Human 2nd round input
<s>bot
Bot 2nd round output<｜end▁of▁sentence｜>
...
...
...
<s>human
Human nth round input
<s>bot
"""

Single-Round without System Prompt:

"""
<s>human
User prompt...
<s>bot

"""

In this format, the system section is optional and the conversation can be either single-turn or multi-turn. When applying inference, you always make your input string end with "<s>bot\n" to ask the model generating answers.

For example, the format used to infer HumanEval is like the following:

<s>human
# language: Python
from typing import List
def separate_paren_groups(paren_string: str) -> List[str]:
    """ Input to this function is a string containing multiple groups of nested parentheses. Your goal is to
    separate those group into separate strings and return the list of those.
    Separate groups are balanced (each open brace is properly closed) and not nested within each other
    Ignore any spaces in the input string.
    >>> separate_paren_groups('( ) (( )) (( )( ))')
    ['()', '(())', '(()())']
    """
<s>bot

Specifically, we also add the Programming Language Tag (e.g. " # language: Python " for Python) used by CodeGeex models.

Quickstart

import os
import torch
import time
from transformers import AutoTokenizer
from auto_gptq import AutoGPTQForCausalLM

os.environ["TOKENIZERS_PARALLELISM"] = "false"

def load_model_tokenizer(model_path):
    """
    Load model and tokenizer based on the given model name or local path of the downloaded model.
    """
    tokenizer = AutoTokenizer.from_pretrained("codefuse-ai/CodeFuse-DeepSeek-33B-4bits", 
                                              trust_remote_code=True, 
                                              use_fast=False,
                                              lagecy=False)
    tokenizer.padding_side = "left"
    tokenizer.pad_token_id = tokenizer.convert_tokens_to_ids("<｜end▁of▁sentence｜>")
    tokenizer.eos_token_id = tokenizer.convert_tokens_to_ids("<｜end▁of▁sentence｜>")

    model = AutoGPTQForCausalLM.from_quantized("codefuse-ai/CodeFuse-DeepSeek-33B-4bits", 
                                                inject_fused_attention=False,
                                                inject_fused_mlp=False,
                                                use_safetensors=True,
                                                use_cuda_fp16=True,
                                                disable_exllama=False,
                                                device_map='auto'   # Support multi-gpus
                                              )
    return model, tokenizer


def inference(model, tokenizer, prompt):
    """
    Uset the given model and tokenizer to generate an answer for the specified prompt.
    """
    st = time.time()
    prompt = prompt if prompt.endswith('\n') else f'{prompt}\n'
    inputs =  f"<s>human\n{prompt}<s>bot\n"

    input_ids = tokenizer.encode(inputs, 
                                  return_tensors="pt", 
                                  padding=True, 
                                  add_special_tokens=False).to("cuda")
    with torch.no_grad():
        generated_ids = model.generate(
            input_ids=input_ids,
            top_p=0.95,
            temperature=0.1,
            do_sample=True,
            max_new_tokens=512,
            eos_token_id=tokenizer.eos_token_id,
            pad_token_id=tokenizer.pad_token_id              
        )
    print(f'generated tokens num is {len(generated_ids[0][input_ids.size(1):])}')
    outputs = tokenizer.batch_decode(generated_ids, skip_special_tokens=True) 
    print(f'generate text is {outputs[0][len(inputs): ]}')
    latency = time.time() - st
    print('latency is {} seconds'.format(latency))

    
if __name__ == "__main__":
    prompt = 'Please write a QuickSort program in Python'

    model, tokenizer = load_model_tokenizer(model_dir)
    inference(model, tokenizer, prompt)

模型简介

CodeFuse-DeepSeek-33B-4bits是代码大模型 CodeFuse-DeepSeek-33B 的4-bits量化版本，后者基于底座模型DeepSeek-Coder-33B使用MFTCoder框架在多个代码相关任务上微调得到。

经过4-bits量化后，CodeFuse-DeepSeek-33B-4bits可在单张A10 (24GB显存)或者RTX 4090（24G显存）上加载。量化后，CodeFuse-DeepSeek-33B-4bits仍取得HumanEval pass@1 78.05%。

新闻

🔥🔥🔥 2024-01-12 CodeFuse-DeepSeek-33B-4bits模型发布。量化后模型在HumanEval pass@1仍取得78.05% (贪婪解码)。

🔥🔥🔥 2024-01-12 CodeFuse-DeepSeek-33B模型发布，模型在HumanEval pass@1指标为78.65% (贪婪解码)。

🔥🔥 2023-11-10 开源了CodeFuse-CodeGeeX2-6B模型，在HumanEval pass@1(greedy decoding)上可以达到48.12%, 比CodeGeeX2提高了9.22%的代码能力（HumanEval）

🔥🔥 2023-10-20 公布了CodeFuse-QWen-14B技术文档，感兴趣详见微信公众号CodeFuse文章： https://mp.weixin.qq.com/s/PCQPkvbvfxSPzsqjOILCDw

🔥🔥 2023-10-16开源了CodeFuse-QWen-14B模型，在HumanEval pass@1(greedy decoding)上可以达到48.78%, 比Qwen-14b提高了16%的代码能力（HumanEval）

🔥🔥 2023-09-27开源了CodeFuse-StarCoder-15B模型，在HumanEval pass@1(greedy decoding)上可以达到54.9%, 比StarCoder提高了21%的代码能力（HumanEval）

🔥🔥🔥 2023-09-26 CodeFuse-CodeLlama-34B 4bits 量化版本发布，量化后模型在HumanEval pass@1指标为73.8% (贪婪解码)。

🔥🔥🔥 2023-09-11 CodeFuse-CodeLlama-34B 发布，HumanEval pass@1指标达到74.4% (贪婪解码), 为当前开源SOTA。

代码社区

大本营 ： 🏡 https://github.com/codefuse-ai （ 请支持我们的项目Star🌟 + Fork🚀 + Watch👀 ）

• 如果您想自己微调该模型，可以访问 ✨ MFTCoder ✨✨

• 如果您想自己部署该模型，可以访问 ✨ FasterTransformer4CodeFuse ✨✨

• 如果您想观看该模型示例，可以访问 ✨ CodeFuse Demo ✨✨

评测表现

代码

Requirements

• python>=3.8
• pytorch>=2.0.0
• transformers>=4.33.2
• Sentencepiece
• auto_gptq
• CUDA 11.4
  

推理数据格式

推理数据为模型在训练数据格式下拼接的字符串形式，它也是推理时输入prompt拼接的方式. 下面分别是带系统提示的多轮会话格式和不带系统提示的单轮会话格式：

带System提示的多轮会话格式:

"""
<s>system
System instruction
<s>human
Human 1st round input
<s>bot
Bot 1st round output<｜end▁of▁sentence｜>
<s>human
Human 2nd round input
<s>bot
Bot 2nd round output<｜end▁of▁sentence｜>
...
...
...
<s>human
Human nth round input
<s>bot
"""

不带System提示的单轮会话格式:

"""
<s>human
User prompt...
<s>bot

"""

在这个格式中，System提示是可选的（按需设定），支持单轮会话也支持多轮会话。推理时，请确保拼接的prompt字符串以"<s>bot\n"结尾，引导模型生成回答。

例如，推理HumanEval数据时使用的格式如下所示：

<s>human
# language: Python
from typing import List
def separate_paren_groups(paren_string: str) -> List[str]:
    """ Input to this function is a string containing multiple groups of nested parentheses. Your goal is to
    separate those group into separate strings and return the list of those.
    Separate groups are balanced (each open brace is properly closed) and not nested within each other
    Ignore any spaces in the input string.
    >>> separate_paren_groups('( ) (( )) (( )( ))')
    ['()', '(())', '(()())']
    """
<s>bot

特别地，我们也使用了CodeGeeX系列模型采用的编程语言区分标签（例如，对于Python语言，我们会使用" # language: Python "）。

快速使用

import os
import torch
import time
from transformers import AutoTokenizer
from auto_gptq import AutoGPTQForCausalLM

os.environ["TOKENIZERS_PARALLELISM"] = "false"

def load_model_tokenizer(model_path):
    """
    Load model and tokenizer based on the given model name or local path of the downloaded model.
    """
    tokenizer = AutoTokenizer.from_pretrained("codefuse-ai/CodeFuse-DeepSeek-33B-4bits", 
                                              trust_remote_code=True, 
                                              use_fast=False,
                                              lagecy=False)
    tokenizer.padding_side = "left"
    tokenizer.pad_token_id = tokenizer.convert_tokens_to_ids("<｜end▁of▁sentence｜>")
    tokenizer.eos_token_id = tokenizer.convert_tokens_to_ids("<｜end▁of▁sentence｜>")

    model = AutoGPTQForCausalLM.from_quantized("codefuse-ai/CodeFuse-DeepSeek-33B-4bits", 
                                                inject_fused_attention=False,
                                                inject_fused_mlp=False,
                                                use_safetensors=True,
                                                use_cuda_fp16=True,
                                                disable_exllama=False,
                                                device_map='auto'   # Support multi-gpus
                                              )
    return model, tokenizer


def inference(model, tokenizer, prompt):
    """
    Uset the given model and tokenizer to generate an answer for the specified prompt.
    """
    st = time.time()
    prompt = prompt if prompt.endswith('\n') else f'{prompt}\n'
    inputs =  f"<s>human\n{prompt}<s>bot\n"

    input_ids = tokenizer.encode(inputs, 
                                  return_tensors="pt", 
                                  padding=True, 
                                  add_special_tokens=False).to("cuda")
    with torch.no_grad():
        generated_ids = model.generate(
            input_ids=input_ids,
            top_p=0.95,
            temperature=0.1,
            do_sample=True,
            max_new_tokens=512,
            eos_token_id=tokenizer.eos_token_id,
            pad_token_id=tokenizer.pad_token_id              
        )
    print(f'generated tokens num is {len(generated_ids[0][input_ids.size(1):])}')
    outputs = tokenizer.batch_decode(generated_ids, skip_special_tokens=True) 
    print(f'generate text is {outputs[0][len(inputs): ]}')
    latency = time.time() - st
    print('latency is {} seconds'.format(latency))

    
if __name__ == "__main__":

    prompt = 'Please write a QuickSort program in Python'

    model, tokenizer = load_model_tokenizer(model_dir)
    inference(model, tokenizer, prompt)