internlm/internlm2_5-7b-chat-1m

InternLM

 

InternLM ^HOT

 

💻Github Repo • 🤔Reporting Issues • 📜Technical Report

👋 join us on Discord and WeChat

Introduction

InternLM2.5 has open-sourced a 7 billion parameter base model and a chat model tailored for practical scenarios. The model has the following characteristics:

• Outstanding reasoning capability: State-of-the-art performance on Math reasoning, surpassing models like Llama3 and Gemma2-9B.

• 1M Context window: Nearly perfect at finding needles in the haystack with 1M-long context, with leading performance on long-context tasks like LongBench. Try it with LMDeploy for 1M-context inference and a file chat demo.

• Stronger tool use: InternLM2.5 supports gathering information from more than 100 web pages, corresponding implementation will be released in Lagent soon. InternLM2.5 has better tool utilization-related capabilities in instruction following, tool selection and reflection. See examples.

InternLM2.5-7B-Chat-1M

InternLM2.5-7B-Chat-1M is the 1M-long-context version of InternLM2.5-7B-Chat.

Performance Evaluation

We employed the "needle in a haystack approach" to evaluate the model's ability to retrieve information from long texts. Results show that InternLM2.5-7B-Chat-1M can accurately locate key information in documents up to 1M tokens in length.

We also used the LongBench benchmark to assess long-document comprehension capabilities. Our model achieved optimal performance in these tests.

LMDeploy

Since huggingface Transformers does not directly support inference with 1M-long context, we recommand to use LMDeploy. The conventional usage with huggingface Transformers is also shown below.

LMDeploy is a toolkit for compressing, deploying, and serving LLM, developed by the MMRazor and MMDeploy teams.

Here is an example of 1M-long context inference. Note: 1M context length requires 4xA100-80G!

pip install lmdeploy

You can run batch inference locally with the following python code:

from lmdeploy import pipeline, GenerationConfig, TurbomindEngineConfig

backend_config = TurbomindEngineConfig(
        rope_scaling_factor=2.5,
        session_len=1048576,  # 1M context length
        max_batch_size=1,
        cache_max_entry_count=0.7,
        tp=4)  # 4xA100-80G.
pipe = pipeline('internlm/internlm2_5-7b-chat-1m', backend_config=backend_config)
prompt = 'Use a long prompt to replace this sentence'
response = pipe(prompt)
print(response)

Find more details in the LMDeploy documentation

Import from Transformers

Since Transformers does not support 1M long context, we only show the usage of non-long context. To load the InternLM2 7B Chat model using Transformers, use the following code:

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("internlm/internlm2_5-7b-chat-1m", trust_remote_code=True)
# Set `torch_dtype=torch.float16` to load model in float16, otherwise it will be loaded as float32 and cause OOM Error.
model = AutoModelForCausalLM.from_pretrained("internlm/internlm2_5-7b-chat-1m", torch_dtype=torch.float16, trust_remote_code=True).cuda()
model = model.eval()
response, history = model.chat(tokenizer, "hello", history=[])
print(response)
# Hello! How can I help you today?
response, history = model.chat(tokenizer, "please provide three suggestions about time management", history=history)
print(response)

The responses can be streamed using stream_chat:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = "internlm/internlm2_5-7b-chat-1m"
model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype=torch.float16, trust_remote_code=True).cuda()
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

model = model.eval()
length = 0
for response, history in model.stream_chat(tokenizer, "Hello", history=[]):
    print(response[length:], flush=True, end="")
    length = len(response)

vLLM

Launch OpenAI compatible server with vLLM>=0.3.2:

pip install vllm

python -m vllm.entrypoints.openai.api_server --model internlm/internlm2_5-7b-chat-1m --served-model-name internlm2_5-7b-chat-1m --trust-remote-code

If you encounter OOM, try to reduce --max-model-len or increase --tensor-parallel-size.

Then you can send a chat request to the server:

curl http://localhost:8000/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
    "model": "internlm2_5-7b-chat-1m",
    "messages": [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Introduce deep learning to me."}
    ]
    }'

Find more details in the vLLM documentation

Open Source License

The code is licensed under Apache-2.0, while model weights are fully open for academic research and also allow free commercial usage. To apply for a commercial license, please fill in the application form (English)/申请表（中文）. For other questions or collaborations, please contact internlm@pjlab.org.cn.

Citation

@misc{cai2024internlm2,
      title={InternLM2 Technical Report},
      author={Zheng Cai and Maosong Cao and Haojiong Chen and Kai Chen and Keyu Chen and Xin Chen and Xun Chen and Zehui Chen and Zhi Chen and Pei Chu and Xiaoyi Dong and Haodong Duan and Qi Fan and Zhaoye Fei and Yang Gao and Jiaye Ge and Chenya Gu and Yuzhe Gu and Tao Gui and Aijia Guo and Qipeng Guo and Conghui He and Yingfan Hu and Ting Huang and Tao Jiang and Penglong Jiao and Zhenjiang Jin and Zhikai Lei and Jiaxing Li and Jingwen Li and Linyang Li and Shuaibin Li and Wei Li and Yining Li and Hongwei Liu and Jiangning Liu and Jiawei Hong and Kaiwen Liu and Kuikun Liu and Xiaoran Liu and Chengqi Lv and Haijun Lv and Kai Lv and Li Ma and Runyuan Ma and Zerun Ma and Wenchang Ning and Linke Ouyang and Jiantao Qiu and Yuan Qu and Fukai Shang and Yunfan Shao and Demin Song and Zifan Song and Zhihao Sui and Peng Sun and Yu Sun and Huanze Tang and Bin Wang and Guoteng Wang and Jiaqi Wang and Jiayu Wang and Rui Wang and Yudong Wang and Ziyi Wang and Xingjian Wei and Qizhen Weng and Fan Wu and Yingtong Xiong and Chao Xu and Ruiliang Xu and Hang Yan and Yirong Yan and Xiaogui Yang and Haochen Ye and Huaiyuan Ying and Jia Yu and Jing Yu and Yuhang Zang and Chuyu Zhang and Li Zhang and Pan Zhang and Peng Zhang and Ruijie Zhang and Shuo Zhang and Songyang Zhang and Wenjian Zhang and Wenwei Zhang and Xingcheng Zhang and Xinyue Zhang and Hui Zhao and Qian Zhao and Xiaomeng Zhao and Fengzhe Zhou and Zaida Zhou and Jingming Zhuo and Yicheng Zou and Xipeng Qiu and Yu Qiao and Dahua Lin},
      year={2024},
      eprint={2403.17297},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

简介

InternLM2.5 ，即书生·浦语大模型第 2.5 代，开源了面向实用场景的70亿参数基础模型与对话模型 （InternLM2.5-7B-Chat）。模型具有以下特点：

• 卓越的推理性能：在数学推理方面取得了同量级模型最优精度，超越了 Llama3 和 Gemma2-9B。
• 有效支持百万字超长上下文：模型在 1 百万字长输入中几乎完美地实现长文“大海捞针”，而且在 LongBench 等长文任务中的表现也达到开源模型中的领先水平。 可以通过 LMDeploy 尝试百万字超长上下文推理。
• 工具调用能力整体升级：InternLM2.5 支持从上百个网页搜集有效信息进行分析推理，相关实现将于近期开源到 Lagent。InternLM2.5 具有更强和更具有泛化性的指令理解、工具筛选与结果反思等能力，新版模型可以更可靠地支持复杂智能体的搭建，支持对工具进行有效的多轮调用，完成较复杂的任务。可以查看更多样例。

InternLM2.5-7B-Chat-1M

InternLM2.5-7B-Chat-1M 支持 1 百万字超长上下文推理，且性能和 InternLM2.5-7B-Chat 相当。考虑到 huggingface Transformers 不直接支持 1M 上下文推理，我们优先推荐使用 lmdeploy 进行百万字超长上下文推理演示。在非超长上下文推理的情况下，你仍然可以使用 huggingface transformers，参考下面的样例代码。

LMDeploy

LMDeploy 由 MMDeploy 和 MMRazor 团队联合开发，是涵盖了 LLM 任务的全套轻量化、部署和服务解决方案。

以下是一个 1M 上下文推理的例子. 注意: 1M 上下文需要 4xA100-80G!

pip install lmdeploy

你可以使用以下 python 代码进行本地批量推理:

from lmdeploy import pipeline, GenerationConfig, TurbomindEngineConfig

backend_config = TurbomindEngineConfig(
        rope_scaling_factor=2.5,
        session_len=1048576,  # 1M context length
        max_batch_size=1,
        cache_max_entry_count=0.7,
        tp=4)  # 4xA100-80G.
pipe = pipeline('internlm/internlm2_5-7b-chat-1m', backend_config=backend_config)
prompt = 'Use a long prompt to replace this sentence'
response = pipe(prompt)
print(response)

通过 Transformers 加载

由于 Transformers 无法支持 1M 长上下文推理，这里仅演示非长文本的用法。

通过以下的代码加载 InternLM2.5 7B Chat 1M 模型

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("internlm/internlm2_5-7b-chat-1m", trust_remote_code=True)
# `torch_dtype=torch.float16` 可以令模型以 float16 精度加载，否则 transformers 会将模型加载为 float32，导致显存不足
model = AutoModelForCausalLM.from_pretrained("internlm/internlm2_5-7b-chat-1m", torch_dtype=torch.float16, trust_remote_code=True).cuda()
model = model.eval()
response, history = model.chat(tokenizer, "你好", history=[])
print(response)
# 你好！有什么我可以帮助你的吗？
response, history = model.chat(tokenizer, "请提供三个管理时间的建议。", history=history)
print(response)

如果想进行流式生成，则可以使用 stream_chat 接口：

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = "internlm/internlm2_5-7b-chat-1m"
model = AutoModelForCausalLM.from_pretrained(model_path, torch_dype=torch.float16, trust_remote_code=True).cuda()
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

model = model.eval()
length = 0
for response, history in model.stream_chat(tokenizer, "你好", history=[]):
    print(response[length:], flush=True, end="")
    length = len(response)

vLLM

使用vLLM>=0.3.2启动兼容 OpenAI API 的服务:

pip install vllm

python -m vllm.entrypoints.openai.api_server --model internlm/internlm2_5-7b-chat-1m --trust-remote-code

如果你遇到 OOM, 请减小 --max-model-len 或增加 --tensor-parallel-size 参数.

然后你可以向服务端发起一个聊天请求:

curl http://localhost:8000/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
    "model": "internlm2_5-7b-chat-1m",
    "messages": [
    {"role": "system", "content": "你是个友善的AI助手。"},
    {"role": "user", "content": "介绍一下深度学习。"}
    ]
    }'

更多信息请查看 vLLM 文档

开源许可证

本仓库的代码依照 Apache-2.0 协议开源。模型权重对学术研究完全开放，也可申请免费的商业使用授权（申请表）。其他问题与合作请联系 internlm@pjlab.org.cn。

引用

@misc{cai2024internlm2,
      title={InternLM2 Technical Report},
      author={Zheng Cai and Maosong Cao and Haojiong Chen and Kai Chen and Keyu Chen and Xin Chen and Xun Chen and Zehui Chen and Zhi Chen and Pei Chu and Xiaoyi Dong and Haodong Duan and Qi Fan and Zhaoye Fei and Yang Gao and Jiaye Ge and Chenya Gu and Yuzhe Gu and Tao Gui and Aijia Guo and Qipeng Guo and Conghui He and Yingfan Hu and Ting Huang and Tao Jiang and Penglong Jiao and Zhenjiang Jin and Zhikai Lei and Jiaxing Li and Jingwen Li and Linyang Li and Shuaibin Li and Wei Li and Yining Li and Hongwei Liu and Jiangning Liu and Jiawei Hong and Kaiwen Liu and Kuikun Liu and Xiaoran Liu and Chengqi Lv and Haijun Lv and Kai Lv and Li Ma and Runyuan Ma and Zerun Ma and Wenchang Ning and Linke Ouyang and Jiantao Qiu and Yuan Qu and Fukai Shang and Yunfan Shao and Demin Song and Zifan Song and Zhihao Sui and Peng Sun and Yu Sun and Huanze Tang and Bin Wang and Guoteng Wang and Jiaqi Wang and Jiayu Wang and Rui Wang and Yudong Wang and Ziyi Wang and Xingjian Wei and Qizhen Weng and Fan Wu and Yingtong Xiong and Chao Xu and Ruiliang Xu and Hang Yan and Yirong Yan and Xiaogui Yang and Haochen Ye and Huaiyuan Ying and Jia Yu and Jing Yu and Yuhang Zang and Chuyu Zhang and Li Zhang and Pan Zhang and Peng Zhang and Ruijie Zhang and Shuo Zhang and Songyang Zhang and Wenjian Zhang and Wenwei Zhang and Xingcheng Zhang and Xinyue Zhang and Hui Zhao and Qian Zhao and Xiaomeng Zhao and Fengzhe Zhou and Zaida Zhou and Jingming Zhuo and Yicheng Zou and Xipeng Qiu and Yu Qiao and Dahua Lin},
      year={2024},
      eprint={2403.17297},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}