InternLM

๐Ÿ‘‹ join us on Discord and WeChat

Introduction

InternLM3 has open-sourced an 8-billion parameter instruction model, InternLM3-8B-Instruct, designed for general-purpose usage and advanced reasoning. This model has the following characteristics:

  • Enhanced performance at reduced cost: State-of-the-art performance on reasoning and knowledge-intensive tasks surpass models like Llama3.1-8B and Qwen2.5-7B. Remarkably, InternLM3 is trained on only 4 trillion high-quality tokens, saving more than 75% of the training cost compared to other LLMs of similar scale.
  • Deep thinking capability: InternLM3 supports both the deep thinking mode for solving complicated reasoning tasks via the long chain-of-thought and the normal response mode for fluent user interactions.

InternLM3-8B-Instruct

Performance Evaluation

We conducted a comprehensive evaluation of InternLM using the open-source evaluation tool OpenCompass. The evaluation covered five dimensions of capabilities: disciplinary competence, language competence, knowledge competence, inference competence, and comprehension competence. Here are some of the evaluation results, and you can visit the OpenCompass leaderboard for more evaluation results.

Benchmark InternLM3-8B-Instruct Qwen2.5-7B-Instruct Llama3.1-8B-Instruct GPT-4o-mini(closed source)
General CMMLU(0-shot) 83.1 75.8 53.9 66.0
MMLU(0-shot) 76.6 76.8 71.8 82.7
MMLU-Pro(0-shot) 57.6 56.2 48.1 64.1
Reasoning GPQA-Diamond(0-shot) 37.4 33.3 24.2 42.9
DROP(0-shot) 83.1 80.4 81.6 85.2
HellaSwag(10-shot) 91.2 85.3 76.7 89.5
KOR-Bench(0-shot) 56.4 44.6 47.7 58.2
MATH MATH-500(0-shot) 83.0* 72.4 48.4 74.0
AIME2024(0-shot) 20.0* 16.7 6.7 13.3
Coding LiveCodeBench(2407-2409 Pass@1) 17.8 16.8 12.9 21.8
HumanEval(Pass@1) 82.3 85.4 72.0 86.6
Instrunction IFEval(Prompt-Strict) 79.3 71.7 75.2 79.7
Long Context RULER(4-128K Average) 87.9 81.4 88.5 90.7
Chat AlpacaEval 2.0(LC WinRate) 51.1 30.3 25.0 50.7
WildBench(Raw Score) 33.1 23.3 1.5 40.3
MT-Bench-101(Score 1-10) 8.59 8.49 8.37 8.87
  • Values marked in bold indicate the highest in open source models
  • The evaluation results were obtained from OpenCompass (some data marked with *, which means evaluating with Thinking Mode), and evaluation configuration can be found in the configuration files provided by OpenCompass.
  • The evaluation data may have numerical differences due to the version iteration of OpenCompass, so please refer to the latest evaluation results of OpenCompass.

Limitations: Although we have made efforts to ensure the safety of the model during the training process and to encourage the model to generate text that complies with ethical and legal requirements, the model may still produce unexpected outputs due to its size and probabilistic generation paradigm. For example, the generated responses may contain biases, discrimination, or other harmful content. Please do not propagate such content. We are not responsible for any consequences resulting from the dissemination of harmful information.

Requirements

transformers >= 4.48

Conversation Mode

Transformers inference

To load the InternLM3 8B Instruct model using Transformers, use the following code:

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_dir = "internlm/internlm3-8b-instruct"
tokenizer = AutoTokenizer.from_pretrained(model_dir, trust_remote_code=True)
# Set `torch_dtype=torch.float16` to load model in float16, otherwise it will be loaded as float32 and might cause OOM Error.
model = AutoModelForCausalLM.from_pretrained(model_dir, trust_remote_code=True, torch_dtype=torch.bfloat16).cuda()
# (Optional) If on low resource devices, you can load model in 4-bit or 8-bit to further save GPU memory via bitsandbytes.
  # InternLM3 8B in 4bit will cost nearly 8GB GPU memory.
  # pip install -U bitsandbytes
  # 8-bit: model = AutoModelForCausalLM.from_pretrained(model_dir, device_map="auto", trust_remote_code=True, load_in_8bit=True)
  # 4-bit: model = AutoModelForCausalLM.from_pretrained(model_dir, device_map="auto", trust_remote_code=True, load_in_4bit=True)
model = model.eval()

system_prompt = """You are an AI assistant whose name is InternLM (ไนฆ็”Ÿยทๆตฆ่ฏญ).
- InternLM (ไนฆ็”Ÿยทๆตฆ่ฏญ) is a conversational language model that is developed by Shanghai AI Laboratory (ไธŠๆตทไบบๅทฅๆ™บ่ƒฝๅฎž้ชŒๅฎค). It is designed to be helpful, honest, and harmless.
- InternLM (ไนฆ็”Ÿยทๆตฆ่ฏญ) can understand and communicate fluently in the language chosen by the user such as English and ไธญๆ–‡."""
messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": "Please tell me five scenic spots in Shanghai"},
 ]
tokenized_chat = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")

generated_ids = model.generate(tokenized_chat, max_new_tokens=1024, temperature=1, repetition_penalty=1.005, top_k=40, top_p=0.8)

generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(tokenized_chat, generated_ids)
]
prompt = tokenizer.batch_decode(tokenized_chat)[0]
print(prompt)
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

LMDeploy inference

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

pip install lmdeploy

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

import lmdeploy
model_dir = "internlm/internlm3-8b-instruct"
pipe = lmdeploy.pipeline(model_dir)
response = pipe("Please tell me five scenic spots in Shanghai")
print(response)

Or you can launch an OpenAI compatible server with the following command:

lmdeploy serve api_server internlm/internlm3-8b-instruct --model-name internlm3-8b-instruct --server-port 23333 

Then you can send a chat request to the server:

curl http://localhost:23333/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
    "model": "internlm3-8b-instruct",
    "messages": [
    {"role": "user", "content": "Please tell me five scenic spots in Shanghai"}
    ]
    }'

Find more details in the LMDeploy documentation

Ollama inference

First install ollama,

# install ollama
curl -fsSL https://ollama.com/install.sh | sh
# fetch model
ollama pull internlm/internlm3-8b-instruct
# install 
pip install ollama

inference code,

import ollama

system_prompt = """You are an AI assistant whose name is InternLM (ไนฆ็”Ÿยทๆตฆ่ฏญ).
- InternLM (ไนฆ็”Ÿยทๆตฆ่ฏญ) is a conversational language model that is developed by Shanghai AI Laboratory (ไธŠๆตทไบบๅทฅๆ™บ่ƒฝๅฎž้ชŒๅฎค). It is designed to be helpful, honest, and harmless.
- InternLM (ไนฆ็”Ÿยทๆตฆ่ฏญ) can understand and communicate fluently in the language chosen by the user such as English and ไธญๆ–‡."""

messages = [
    {
        "role": "system",
        "content": system_prompt,
    },
    {
        "role": "user",
        "content": "Please tell me five scenic spots in Shanghai"
    },
]

stream = ollama.chat(
    model='internlm/internlm3-8b-instruct',
    messages=messages,
    stream=True,
)

for chunk in stream:
  print(chunk['message']['content'], end='', flush=True)

vLLM inference

Refer to installation to install the latest code of vllm

pip install vllm --pre --extra-index-url https://wheels.vllm.ai/nightly

inference code:

from vllm import LLM, SamplingParams

llm = LLM(model="internlm/internlm3-8b-instruct")
sampling_params = SamplingParams(temperature=1, repetition_penalty=1.005, top_k=40, top_p=0.8)

system_prompt = """You are an AI assistant whose name is InternLM (ไนฆ็”Ÿยทๆตฆ่ฏญ).
- InternLM (ไนฆ็”Ÿยทๆตฆ่ฏญ) is a conversational language model that is developed by Shanghai AI Laboratory (ไธŠๆตทไบบๅทฅๆ™บ่ƒฝๅฎž้ชŒๅฎค). It is designed to be helpful, honest, and harmless.
- InternLM (ไนฆ็”Ÿยทๆตฆ่ฏญ) can understand and communicate fluently in the language chosen by the user such as English and ไธญๆ–‡."""

prompts = [
    {
        "role": "system",
        "content": system_prompt,
    },
    {
        "role": "user",
        "content": "Please tell me five scenic spots in Shanghai"
    },
]
outputs = llm.chat(prompts,
                   sampling_params=sampling_params,
                   use_tqdm=False)
print(outputs)

Thinking Mode

Thinking Demo

Thinking system prompt

thinking_system_prompt = """You are an expert mathematician with extensive experience in mathematical competitions. You approach problems through systematic thinking and rigorous reasoning. When solving problems, follow these thought processes:
## Deep Understanding
Take time to fully comprehend the problem before attempting a solution. Consider:
- What is the real question being asked?
- What are the given conditions and what do they tell us?
- Are there any special restrictions or assumptions?
- Which information is crucial and which is supplementary?
## Multi-angle Analysis
Before solving, conduct thorough analysis:
- What mathematical concepts and properties are involved?
- Can you recall similar classic problems or solution methods?
- Would diagrams or tables help visualize the problem?
- Are there special cases that need separate consideration?
## Systematic Thinking
Plan your solution path:
- Propose multiple possible approaches
- Analyze the feasibility and merits of each method
- Choose the most appropriate method and explain why
- Break complex problems into smaller, manageable steps
## Rigorous Proof
During the solution process:
- Provide solid justification for each step
- Include detailed proofs for key conclusions
- Pay attention to logical connections
- Be vigilant about potential oversights
## Repeated Verification
After completing your solution:
- Verify your results satisfy all conditions
- Check for overlooked special cases
- Consider if the solution can be optimized or simplified
- Review your reasoning process
Remember:
1. Take time to think thoroughly rather than rushing to an answer
2. Rigorously prove each key conclusion
3. Keep an open mind and try different approaches
4. Summarize valuable problem-solving methods
5. Maintain healthy skepticism and verify multiple times
Your response should reflect deep mathematical understanding and precise logical thinking, making your solution path and reasoning clear to others.
When you're ready, present your complete solution with:
- Clear problem understanding
- Detailed solution process
- Key insights
- Thorough verification
Focus on clear, logical progression of ideas and thorough explanation of your mathematical reasoning. Provide answers in the same language as the user asking the question, repeat the final answer using a '\\boxed{}' without any units, you have [[8192]] tokens to complete the answer.
"""

Transformers inference

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_dir = "internlm/internlm3-8b-instruct"
tokenizer = AutoTokenizer.from_pretrained(model_dir, trust_remote_code=True)
# Set `torch_dtype=torch.float16` to load model in float16, otherwise it will be loaded as float32 and might cause OOM Error.
model = AutoModelForCausalLM.from_pretrained(model_dir, trust_remote_code=True, torch_dtype=torch.bfloat16).cuda()
# (Optional) If on low resource devices, you can load model in 4-bit or 8-bit to further save GPU memory via bitsandbytes.
  # InternLM3 8B in 4bit will cost nearly 8GB GPU memory.
  # pip install -U bitsandbytes
  # 8-bit: model = AutoModelForCausalLM.from_pretrained(model_dir, device_map="auto", trust_remote_code=True, load_in_8bit=True)
  # 4-bit: model = AutoModelForCausalLM.from_pretrained(model_dir, device_map="auto", trust_remote_code=True, load_in_4bit=True)
model = model.eval()

messages = [
    {"role": "system", "content": thinking_system_prompt},
    {"role": "user", "content": "Given the function\(f(x)=\mathrm{e}^{x}-ax - a^{3}\),\n(1) When \(a = 1\), find the equation of the tangent line to the curve \(y = f(x)\) at the point \((1,f(1))\).\n(2) If \(f(x)\) has a local minimum and the minimum value is less than \(0\), determine the range of values for \(a\)."},
 ]
tokenized_chat = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")

generated_ids = model.generate(tokenized_chat, max_new_tokens=8192)

generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(tokenized_chat, generated_ids)
]
prompt = tokenizer.batch_decode(tokenized_chat)[0]
print(prompt)
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

LMDeploy inference

LMDeploy is a toolkit for compressing, deploying, and serving LLM.

pip install lmdeploy

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

from lmdeploy import pipeline, GenerationConfig, ChatTemplateConfig
model_dir = "internlm/internlm3-8b-instruct"
chat_template_config = ChatTemplateConfig(model_name='internlm3')
pipe = pipeline(model_dir, chat_template_config=chat_template_config)

messages = [
        {"role": "system", "content": thinking_system_prompt},
        {"role": "user", "content": "Given the function\(f(x)=\mathrm{e}^{x}-ax - a^{3}\),\n(1) When \(a = 1\), find the equation of the tangent line to the curve \(y = f(x)\) at the point \((1,f(1))\).\n(2) If \(f(x)\) has a local minimum and the minimum value is less than \(0\), determine the range of values for \(a\)."},
]

response = pipe(messages, gen_config=GenerationConfig(max_new_tokens=2048))
print(response)

Ollama inference

First install ollama,

# install ollama
curl -fsSL https://ollama.com/install.sh | sh
# fetch model
ollama pull internlm/internlm3-8b-instruct
# install
pip install ollama

inference code,

import ollama

messages = [
    {
        "role": "system",
        "content": thinking_system_prompt,
    },
    {
        "role": "user",
        "content": "Given the function\(f(x)=\mathrm{e}^{x}-ax - a^{3}\),\n(1) When \(a = 1\), find the equation of the tangent line to the curve \(y = f(x)\) at the point \((1,f(1))\).\n(2) If \(f(x)\) has a local minimum and the minimum value is less than \(0\), determine the range of values for \(a\)."
    },
]

stream = ollama.chat(
    model='internlm/internlm3-8b-instruct',
    messages=messages,
    stream=True,
)

for chunk in stream:
  print(chunk['message']['content'], end='', flush=True)

vLLM inference

Refer to installation to install the latest code of vllm

pip install vllm --pre --extra-index-url https://wheels.vllm.ai/nightly

inference code

from vllm import LLM, SamplingParams

llm = LLM(model="internlm/internlm3-8b-instruct")
sampling_params = SamplingParams(temperature=1, repetition_penalty=1.005, top_k=40, top_p=0.8, max_tokens=8192)

prompts = [
    {
        "role": "system",
        "content": thinking_system_prompt,
    },
    {
        "role": "user",
        "content": "Given the function\(f(x)=\mathrm{e}^{x}-ax - a^{3}\),\n(1) When \(a = 1\), find the equation of the tangent line to the curve \(y = f(x)\) at the point \((1,f(1))\).\n(2) If \(f(x)\) has a local minimum and the minimum value is less than \(0\), determine the range of values for \(a\)."
    },
]
outputs = llm.chat(prompts,
                   sampling_params=sampling_params,
                   use_tqdm=False)
print(outputs)

Open Source License

Code and model weights are licensed under Apache-2.0.

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}
}

็ฎ€ไป‹

InternLM3-8B-Instruct

InternLM3๏ผŒๅณไนฆ็”Ÿยทๆตฆ่ฏญๅคงๆจกๅž‹็ฌฌ3ไปฃ๏ผŒๅผ€ๆบไบ†80ไบฟๅ‚ๆ•ฐ๏ผŒ้ขๅ‘้€š็”จไฝฟ็”จไธŽ้ซ˜้˜ถๆŽจ็†็š„ๆŒ‡ไปคๆจกๅž‹๏ผˆInternLM3-8B-Instruct๏ผ‰ใ€‚ๆจกๅž‹ๅ…ทๅค‡ไปฅไธ‹็‰น็‚น๏ผš

  • ๆ›ดไฝŽ็š„ไปฃไปทๅ–ๅพ—ๆ›ด้ซ˜็š„ๆ€ง่ƒฝ: ๅœจๆŽจ็†ใ€็Ÿฅ่ฏ†็ฑปไปปๅŠกไธŠๅ–ๅพ—ๅŒ้‡็บงๆœ€ไผ˜ๆ€ง่ƒฝ๏ผŒ่ถ…่ฟ‡Llama3.1-8Bๅ’ŒQwen2.5-7Bใ€‚ๅ€ผๅพ—ๅ…ณๆณจ็š„ๆ˜ฏInternLM3ๅช็”จไบ†4ไธ‡ไบฟ่ฏๅ…ƒ่ฟ›่กŒ่ฎญ็ปƒ๏ผŒๅฏนๆฏ”ๅŒ็บงๅˆซๆจกๅž‹่ฎญ็ปƒๆˆๆœฌ่Š‚็œ75%ไปฅไธŠใ€‚
  • ๆทฑๅบฆๆ€่€ƒ่ƒฝๅŠ›: InternLM3ๆ”ฏๆŒ้€š่ฟ‡้•ฟๆ€็ปด้“พๆฑ‚่งฃๅคๆ‚ๆŽจ็†ไปปๅŠก็š„ๆทฑๅบฆๆ€่€ƒๆจกๅผ๏ผŒๅŒๆ—ถ่ฟ˜ๅ…ผ้กพไบ†็”จๆˆทไฝ“้ชŒๆ›ดๆต็•…็š„้€š็”จๅ›žๅคๆจกๅผใ€‚

ๆ€ง่ƒฝ่ฏ„ๆต‹

ๆˆ‘ไปฌไฝฟ็”จๅผ€ๆบ่ฏ„ๆต‹ๅทฅๅ…ท OpenCompass ไปŽๅญฆ็ง‘็ปผๅˆ่ƒฝๅŠ›ใ€่ฏญ่จ€่ƒฝๅŠ›ใ€็Ÿฅ่ฏ†่ƒฝๅŠ›ใ€ๆŽจ็†่ƒฝๅŠ›ใ€็†่งฃ่ƒฝๅŠ›ไบ”ๅคง่ƒฝๅŠ›็ปดๅบฆๅฏนInternLMๅผ€ๅฑ•ๅ…จ้ข่ฏ„ๆต‹๏ผŒ้ƒจๅˆ†่ฏ„ๆต‹็ป“ๆžœๅฆ‚ไธ‹่กจๆ‰€็คบ๏ผŒๆฌข่ฟŽ่ฎฟ้—ฎ OpenCompass ๆฆœๅ• ่Žทๅ–ๆ›ดๅคš็š„่ฏ„ๆต‹็ป“ๆžœใ€‚

่ฏ„ๆต‹้›†\ๆจกๅž‹ InternLM3-8B-Instruct Qwen2.5-7B-Instruct Llama3.1-8B-Instruct GPT-4o-mini(้—ญๆบ)
General CMMLU(0-shot) 83.1 75.8 53.9 66.0
MMLU(0-shot) 76.6 76.8 71.8 82.7
MMLU-Pro(0-shot) 57.6 56.2 48.1 64.1
Reasoning GPQA-Diamond(0-shot) 37.4 33.3 24.2 42.9
DROP(0-shot) 83.1 80.4 81.6 85.2
HellaSwag(10-shot) 91.2 85.3 76.7 89.5
KOR-Bench(0-shot) 56.4 44.6 47.7 58.2
MATH MATH-500(0-shot) 83.0* 72.4 48.4 74.0
AIME2024(0-shot) 20.0* 16.7 6.7 13.3
Coding LiveCodeBench(2407-2409 Pass@1) 17.8 16.8 12.9 21.8
HumanEval(Pass@1) 82.3 85.4 72.0 86.6
Instrunction IFEval(Prompt-Strict) 79.3 71.7 75.2 79.7
LongContext RULER(4-128K Average) 87.9 81.4 88.5 90.7
Chat AlpacaEval 2.0(LC WinRate) 51.1 30.3 25.0 50.7
WildBench(Raw Score) 33.1 23.3 1.5 40.3
MT-Bench-101(Score 1-10) 8.59 8.49 8.37 8.87
  • ่กจไธญๆ ‡็ฒ—็š„ๆ•ฐๅ€ผ่กจ็คบๅœจๅฏนๆฏ”็š„ๅผ€ๆบๆจกๅž‹ไธญ็š„ๆœ€้ซ˜ๅ€ผใ€‚
  • ไปฅไธŠ่ฏ„ๆต‹็ป“ๆžœๅŸบไบŽ OpenCompass ่Žทๅพ—(้ƒจๅˆ†ๆ•ฐๆฎๆ ‡ๆณจ*ไปฃ่กจไฝฟ็”จๆทฑๅบฆๆ€่€ƒๆจกๅผ่ฟ›่กŒ่ฏ„ๆต‹)๏ผŒๅ…ทไฝ“ๆต‹่ฏ•็ป†่Š‚ๅฏๅ‚่ง OpenCompass ไธญๆไพ›็š„้…็ฝฎๆ–‡ไปถใ€‚
  • ่ฏ„ๆต‹ๆ•ฐๆฎไผšๅ›  OpenCompass ็š„็‰ˆๆœฌ่ฟญไปฃ่€Œๅญ˜ๅœจๆ•ฐๅ€ผๅทฎๅผ‚๏ผŒ่ฏทไปฅ OpenCompass ๆœ€ๆ–ฐ็‰ˆ็š„่ฏ„ๆต‹็ป“ๆžœไธบไธปใ€‚

ๅฑ€้™ๆ€ง๏ผš ๅฐฝ็ฎกๅœจ่ฎญ็ปƒ่ฟ‡็จ‹ไธญๆˆ‘ไปฌ้žๅธธๆณจ้‡ๆจกๅž‹็š„ๅฎ‰ๅ…จๆ€ง๏ผŒๅฐฝๅŠ›ไฟƒไฝฟๆจกๅž‹่พ“ๅ‡บ็ฌฆๅˆไผฆ็†ๅ’Œๆณ•ๅพ‹่ฆๆฑ‚็š„ๆ–‡ๆœฌ๏ผŒไฝ†ๅ—้™ไบŽๆจกๅž‹ๅคงๅฐไปฅๅŠๆฆ‚็އ็”Ÿๆˆ่Œƒๅผ๏ผŒๆจกๅž‹ๅฏ่ƒฝไผšไบง็”Ÿๅ„็งไธ็ฌฆๅˆ้ข„ๆœŸ็š„่พ“ๅ‡บ๏ผŒไพ‹ๅฆ‚ๅ›žๅคๅ†…ๅฎนๅŒ…ๅซๅ่งใ€ๆญง่ง†็ญ‰ๆœ‰ๅฎณๅ†…ๅฎน๏ผŒ่ฏทๅ‹ฟไผ ๆ’ญ่ฟ™ไบ›ๅ†…ๅฎนใ€‚็”ฑไบŽไผ ๆ’ญไธ่‰ฏไฟกๆฏๅฏผ่‡ด็š„ไปปไฝ•ๅŽๆžœ๏ผŒๆœฌ้กน็›ฎไธๆ‰ฟๆ‹…่ดฃไปปใ€‚

ไพ่ต–

transformers >= 4.48

ๅธธ่ง„ๅฏน่ฏๆจกๅผ

Transformers ๆŽจ็†

้€š่ฟ‡ไปฅไธ‹็š„ไปฃ็ ๅŠ ่ฝฝ InternLM3 8B Instruct ๆจกๅž‹

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_dir = "internlm/internlm3-8b-instruct"
tokenizer = AutoTokenizer.from_pretrained(model_dir, trust_remote_code=True)
# Set `torch_dtype=torch.float16` to load model in float16, otherwise it will be loaded as float32 and might cause OOM Error.
model = AutoModelForCausalLM.from_pretrained(model_dir, trust_remote_code=True, torch_dtype=torch.bfloat16).cuda()
# (Optional) If on low resource devices, you can load model in 4-bit or 8-bit to further save GPU memory via bitsandbytes.
  # InternLM3 8B in 4bit will cost nearly 8GB GPU memory.
  # pip install -U bitsandbytes
  # 8-bit: model = AutoModelForCausalLM.from_pretrained(model_dir, device_map="auto", trust_remote_code=True, load_in_8bit=True)
  # 4-bit: model = AutoModelForCausalLM.from_pretrained(model_dir, device_map="auto", trust_remote_code=True, load_in_4bit=True)
model = model.eval()

system_prompt = """You are an AI assistant whose name is InternLM (ไนฆ็”Ÿยทๆตฆ่ฏญ).
- InternLM (ไนฆ็”Ÿยทๆตฆ่ฏญ) is a conversational language model that is developed by Shanghai AI Laboratory (ไธŠๆตทไบบๅทฅๆ™บ่ƒฝๅฎž้ชŒๅฎค). It is designed to be helpful, honest, and harmless.
- InternLM (ไนฆ็”Ÿยทๆตฆ่ฏญ) can understand and communicate fluently in the language chosen by the user such as English and ไธญๆ–‡."""
messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": "Please tell me five scenic spots in Shanghai"},
 ]
tokenized_chat = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")

generated_ids = model.generate(tokenized_chat, max_new_tokens=1024, temperature=1, repetition_penalty=1.005, top_k=40, top_p=0.8)

generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(tokenized_chat, generated_ids)
]
prompt = tokenizer.batch_decode(tokenized_chat)[0]
print(prompt)
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)
LMDeploy ๆŽจ็†

LMDeploy ๆ˜ฏๆถต็›–ไบ† LLM ไปปๅŠก็š„ๅ…จๅฅ—่ฝป้‡ๅŒ–ใ€้ƒจ็ฝฒๅ’ŒๆœๅŠก่งฃๅ†ณๆ–นๆกˆใ€‚

pip install lmdeploy

ไฝ ๅฏไปฅไฝฟ็”จไปฅไธ‹ python ไปฃ็ ่ฟ›่กŒๆœฌๅœฐๆ‰น้‡ๆŽจ็†:

import lmdeploy
model_dir = "internlm/internlm3-8b-instruct"
pipe = lmdeploy.pipeline(model_dir)
response = pipe(["Please tell me five scenic spots in Shanghai"])
print(response)

ๆˆ–่€…ไฝ ๅฏไปฅไฝฟ็”จไปฅไธ‹ๅ‘ฝไปคๅฏๅŠจๅ…ผๅฎน OpenAI API ็š„ๆœๅŠก:

lmdeploy serve api_server internlm/internlm3-8b-instruct --model-name internlm3-8b-instruct --server-port 23333 

็„ถๅŽไฝ ๅฏไปฅๅ‘ๆœๅŠก็ซฏๅ‘่ตทไธ€ไธช่Šๅคฉ่ฏทๆฑ‚:

curl http://localhost:23333/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
    "model": "internlm3-8b-instruct",
    "messages": [
    {"role": "user", "content": "ไป‹็ปไธ€ไธ‹ๆทฑๅบฆๅญฆไน ใ€‚"}
    ]
    }'

ๆ›ดๅคšไฟกๆฏ่ฏทๆŸฅ็œ‹ LMDeploy ๆ–‡ๆกฃ

Ollama ๆŽจ็†

ๅ‡†ๅค‡ๅทฅไฝœ

# install ollama
curl -fsSL https://ollama.com/install.sh | sh
# fetch ๆจกๅž‹
ollama pull internlm/internlm3-8b-instruct
# install pythonๅบ“
pip install ollama

ๆŽจ็†ไปฃ็ 

import ollama

system_prompt = """You are an AI assistant whose name is InternLM (ไนฆ็”Ÿยทๆตฆ่ฏญ).
- InternLM (ไนฆ็”Ÿยทๆตฆ่ฏญ) is a conversational language model that is developed by Shanghai AI Laboratory (ไธŠๆตทไบบๅทฅๆ™บ่ƒฝๅฎž้ชŒๅฎค). It is designed to be helpful, honest, and harmless.
- InternLM (ไนฆ็”Ÿยทๆตฆ่ฏญ) can understand and communicate fluently in the language chosen by the user such as English and ไธญๆ–‡."""

messages = [
    {
        "role": "system",
        "content": system_prompt,
    },
    {
        "role": "user",
        "content": "Please tell me five scenic spots in Shanghai"
    },
]

stream = ollama.chat(
    model='internlm/internlm3-8b-instruct',
    messages=messages,
    stream=True,
)

for chunk in stream:
  print(chunk['message']['content'], end='', flush=True)

vLLM ๆŽจ็†

ๅ‚่€ƒๆ–‡ๆกฃ ๅฎ‰่ฃ… vllm ๆœ€ๆ–ฐไปฃ็ 

pip install vllm --pre --extra-index-url https://wheels.vllm.ai/nightly

ๆŽจ็†ไปฃ็ 

from vllm import LLM, SamplingParams

llm = LLM(model="internlm/internlm3-8b-instruct")
sampling_params = SamplingParams(temperature=1, repetition_penalty=1.005, top_k=40, top_p=0.8)

system_prompt = """You are an AI assistant whose name is InternLM (ไนฆ็”Ÿยทๆตฆ่ฏญ).
- InternLM (ไนฆ็”Ÿยทๆตฆ่ฏญ) is a conversational language model that is developed by Shanghai AI Laboratory (ไธŠๆตทไบบๅทฅๆ™บ่ƒฝๅฎž้ชŒๅฎค). It is designed to be helpful, honest, and harmless.
- InternLM (ไนฆ็”Ÿยทๆตฆ่ฏญ) can understand and communicate fluently in the language chosen by the user such as English and ไธญๆ–‡."""

prompts = [
    {
        "role": "system",
        "content": system_prompt,
    },
    {
        "role": "user",
        "content": "Please tell me five scenic spots in Shanghai"
    },
]
outputs = llm.chat(prompts,
                   sampling_params=sampling_params,
                   use_tqdm=False)
print(outputs)

ๆทฑๅบฆๆ€่€ƒๆจกๅผ

ๆทฑๅบฆๆ€่€ƒ Demo
ๆทฑๅบฆๆ€่€ƒ system prompt
thinking_system_prompt = """You are an expert mathematician with extensive experience in mathematical competitions. You approach problems through systematic thinking and rigorous reasoning. When solving problems, follow these thought processes:
## Deep Understanding
Take time to fully comprehend the problem before attempting a solution. Consider:
- What is the real question being asked?
- What are the given conditions and what do they tell us?
- Are there any special restrictions or assumptions?
- Which information is crucial and which is supplementary?
## Multi-angle Analysis
Before solving, conduct thorough analysis:
- What mathematical concepts and properties are involved?
- Can you recall similar classic problems or solution methods?
- Would diagrams or tables help visualize the problem?
- Are there special cases that need separate consideration?
## Systematic Thinking
Plan your solution path:
- Propose multiple possible approaches
- Analyze the feasibility and merits of each method
- Choose the most appropriate method and explain why
- Break complex problems into smaller, manageable steps
## Rigorous Proof
During the solution process:
- Provide solid justification for each step
- Include detailed proofs for key conclusions
- Pay attention to logical connections
- Be vigilant about potential oversights
## Repeated Verification
After completing your solution:
- Verify your results satisfy all conditions
- Check for overlooked special cases
- Consider if the solution can be optimized or simplified
- Review your reasoning process
Remember:
1. Take time to think thoroughly rather than rushing to an answer
2. Rigorously prove each key conclusion
3. Keep an open mind and try different approaches
4. Summarize valuable problem-solving methods
5. Maintain healthy skepticism and verify multiple times
Your response should reflect deep mathematical understanding and precise logical thinking, making your solution path and reasoning clear to others.
When you're ready, present your complete solution with:
- Clear problem understanding
- Detailed solution process
- Key insights
- Thorough verification
Focus on clear, logical progression of ideas and thorough explanation of your mathematical reasoning. Provide answers in the same language as the user asking the question, repeat the final answer using a '\\boxed{}' without any units, you have [[8192]] tokens to complete the answer.
"""
Transformers ๆŽจ็†
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_dir = "internlm/internlm3-8b-instruct"
tokenizer = AutoTokenizer.from_pretrained(model_dir, trust_remote_code=True)
# Set `torch_dtype=torch.float16` to load model in float16, otherwise it will be loaded as float32 and might cause OOM Error.
model = AutoModelForCausalLM.from_pretrained(model_dir, trust_remote_code=True, torch_dtype=torch.bfloat16).cuda()
# (Optional) If on low resource devices, you can load model in 4-bit or 8-bit to further save GPU memory via bitsandbytes.
  # InternLM3 8B in 4bit will cost nearly 8GB GPU memory.
  # pip install -U bitsandbytes
  # 8-bit: model = AutoModelForCausalLM.from_pretrained(model_dir, device_map="auto", trust_remote_code=True, load_in_8bit=True)
  # 4-bit: model = AutoModelForCausalLM.from_pretrained(model_dir, device_map="auto", trust_remote_code=True, load_in_4bit=True)
model = model.eval()

messages = [
    {"role": "system", "content": thinking_system_prompt},
    {"role": "user", "content": "ๅทฒ็Ÿฅๅ‡ฝๆ•ฐ\(f(x)=\mathrm{e}^{x}-ax - a^{3}\)ใ€‚\n๏ผˆ1๏ผ‰ๅฝ“\(a = 1\)ๆ—ถ๏ผŒๆฑ‚ๆ›ฒ็บฟ\(y = f(x)\)ๅœจ็‚น\((1,f(1))\)ๅค„็š„ๅˆ‡็บฟๆ–น็จ‹๏ผ›\n๏ผˆ2๏ผ‰่‹ฅ\(f(x)\)ๆœ‰ๆžๅฐๅ€ผ๏ผŒไธ”ๆžๅฐๅ€ผๅฐไบŽ\(0\)๏ผŒๆฑ‚\(a\)็š„ๅ–ๅ€ผ่Œƒๅ›ดใ€‚"},
 ]
tokenized_chat = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")

generated_ids = model.generate(tokenized_chat, max_new_tokens=8192)

generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(tokenized_chat, generated_ids)
]
prompt = tokenizer.batch_decode(tokenized_chat)[0]
print(prompt)
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)
LMDeploy ๆŽจ็†

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

pip install lmdeploy

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

from lmdeploy import pipeline, GenerationConfig, ChatTemplateConfig
model_dir = "internlm/internlm3-8b-instruct"
chat_template_config = ChatTemplateConfig(model_name='internlm3')
pipe = pipeline(model_dir, chat_template_config=chat_template_config)

messages = [
        {"role": "system", "content": thinking_system_prompt},
        {"role": "user", "content": "ๅทฒ็Ÿฅๅ‡ฝๆ•ฐ\(f(x)=\mathrm{e}^{x}-ax - a^{3}\)ใ€‚\n๏ผˆ1๏ผ‰ๅฝ“\(a = 1\)ๆ—ถ๏ผŒๆฑ‚ๆ›ฒ็บฟ\(y = f(x)\)ๅœจ็‚น\((1,f(1))\)ๅค„็š„ๅˆ‡็บฟๆ–น็จ‹๏ผ›\n๏ผˆ2๏ผ‰่‹ฅ\(f(x)\)ๆœ‰ๆžๅฐๅ€ผ๏ผŒไธ”ๆžๅฐๅ€ผๅฐไบŽ\(0\)๏ผŒๆฑ‚\(a\)็š„ๅ–ๅ€ผ่Œƒๅ›ดใ€‚"},
]

response = pipe(messages, gen_config=GenerationConfig(max_new_tokens=2048))
print(response)
Ollama ๆŽจ็†

ๅ‡†ๅค‡ๅทฅไฝœ

# install ollama
curl -fsSL https://ollama.com/install.sh | sh
# fetch ๆจกๅž‹
ollama pull internlm/internlm3-8b-instruct
# install pythonๅบ“
pip install ollama

inference code,

import ollama

messages = [
    {
        "role": "system",
        "content": thinking_system_prompt,
    },
    {
        "role": "user",
        "content": "Given the function\(f(x)=\mathrm{e}^{x}-ax - a^{3}\),\n(1) When \(a = 1\), find the equation of the tangent line to the curve \(y = f(x)\) at the point \((1,f(1))\).\n(2) If \(f(x)\) has a local minimum and the minimum value is less than \(0\), determine the range of values for \(a\)."
    },
]

stream = ollama.chat(
    model='internlm/internlm3-8b-instruct',
    messages=messages,
    stream=True,
)

for chunk in stream:
  print(chunk['message']['content'], end='', flush=True)

vLLM ๆŽจ็†

ๅ‚่€ƒๆ–‡ๆกฃ ๅฎ‰่ฃ… vllm ๆœ€ๆ–ฐไปฃ็ 

pip install vllm --pre --extra-index-url https://wheels.vllm.ai/nightly

ๆŽจ็†ไปฃ็ 

from vllm import LLM, SamplingParams

llm = LLM(model="internlm/internlm3-8b-instruct")
sampling_params = SamplingParams(temperature=1, repetition_penalty=1.005, top_k=40, top_p=0.8, max_tokens=8192)

prompts = [
    {
        "role": "system",
        "content": thinking_system_prompt,
    },
    {
        "role": "user",
        "content": "ๅทฒ็Ÿฅๅ‡ฝๆ•ฐ\(f(x)=\mathrm{e}^{x}-ax - a^{3}\)ใ€‚\n๏ผˆ1๏ผ‰ๅฝ“\(a = 1\)ๆ—ถ๏ผŒๆฑ‚ๆ›ฒ็บฟ\(y = f(x)\)ๅœจ็‚น\((1,f(1))\)ๅค„็š„ๅˆ‡็บฟๆ–น็จ‹๏ผ›\n๏ผˆ2๏ผ‰่‹ฅ\(f(x)\)ๆœ‰ๆžๅฐๅ€ผ๏ผŒไธ”ๆžๅฐๅ€ผๅฐไบŽ\(0\)๏ผŒๆฑ‚\(a\)็š„ๅ–ๅ€ผ่Œƒๅ›ดใ€‚"
    },
]
outputs = llm.chat(prompts,
                   sampling_params=sampling_params,
                   use_tqdm=False)
print(outputs)

ๅผ€ๆบ่ฎธๅฏ่ฏ

ๆœฌไป“ๅบ“็š„ไปฃ็ ๅ’Œๆƒ้‡ไพ็…ง Apache-2.0 ๅ่ฎฎๅผ€ๆบใ€‚

ๅผ•็”จ

@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}
}
Downloads last month
18,610
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ 1 Ask for provider support

Model tree for internlm/internlm3-8b-instruct

Adapters
5 models
Finetunes
5 models
Merges
2 models
Quantizations
45 models

Spaces using internlm/internlm3-8b-instruct 39

Collection including internlm/internlm3-8b-instruct

Paper for internlm/internlm3-8b-instruct