ChatGLM.cpp

C++ implementation of ChatGLM-6B, ChatGLM2-6B, ChatGLM3, GLM-4 and more LLMs for real-time chatting on your MacBook.

Features

Highlights:

Pure C++ implementation based on ggml, working in the same way as llama.cpp.
Accelerated memory-efficient CPU inference with int4/int8 quantization, optimized KV cache and parallel computing.
P-Tuning v2 and LoRA finetuned models support.
Streaming generation with typewriter effect.
Python binding, web demo, api servers and more possibilities.

Support Matrix:

Hardwares: x86/arm CPU, NVIDIA GPU, Apple Silicon GPU
Platforms: Linux, MacOS, Windows
Models: ChatGLM-6B, ChatGLM2-6B, ChatGLM3, GLM-4, CodeGeeX2, Baichuan-13B, Baichuan-7B, Baichuan-13B, Baichuan2, InternLM

NOTE: Baichuan & InternLM model series are deprecated in favor of llama.cpp.

Getting Started

Preparation

Clone the ChatGLM.cpp repository into your local machine:

git clone --recursive https://github.com/li-plus/chatglm.cpp.git && cd chatglm.cpp

If you forgot the --recursive flag when cloning the repository, run the following command in the chatglm.cpp folder:

git submodule update --init --recursive

Quantize Model

Install necessary packages for loading and quantizing Hugging Face models:

python3 -m pip install -U pip
python3 -m pip install torch tabulate tqdm transformers accelerate sentencepiece

Use convert.py to transform ChatGLM-6B into quantized GGML format. For example, to convert the fp16 original model to q4_0 (quantized int4) GGML model, run:

python3 chatglm_cpp/convert.py -i THUDM/chatglm-6b -t q4_0 -o models/chatglm-ggml.bin

The original model (-i <model_name_or_path>) can be a Hugging Face model name or a local path to your pre-downloaded model. Currently supported models are:

ChatGLM-6B: THUDM/chatglm-6b, THUDM/chatglm-6b-int8, THUDM/chatglm-6b-int4
ChatGLM2-6B: THUDM/chatglm2-6b, THUDM/chatglm2-6b-int4
ChatGLM3-6B: THUDM/chatglm3-6b
ChatGLM4-9B: THUDM/glm-4-9b-chat
CodeGeeX2: THUDM/codegeex2-6b, THUDM/codegeex2-6b-int4
Baichuan & Baichuan2: baichuan-inc/Baichuan-13B-Chat, baichuan-inc/Baichuan2-7B-Chat, baichuan-inc/Baichuan2-13B-Chat

You are free to try any of the below quantization types by specifying -t <type>:

q4_0: 4-bit integer quantization with fp16 scales.
q4_1: 4-bit integer quantization with fp16 scales and minimum values.
q5_0: 5-bit integer quantization with fp16 scales.
q5_1: 5-bit integer quantization with fp16 scales and minimum values.
q8_0: 8-bit integer quantization with fp16 scales.
f16: half precision floating point weights without quantization.
f32: single precision floating point weights without quantization.

For LoRA models, add -l <lora_model_name_or_path> flag to merge your LoRA weights into the base model. For example, run python3 chatglm_cpp/convert.py -i THUDM/chatglm3-6b -t q4_0 -o models/chatglm3-ggml-lora.bin -l shibing624/chatglm3-6b-csc-chinese-lora to merge public LoRA weights from Hugging Face.

For P-Tuning v2 models using the official finetuning script, additional weights are automatically detected by convert.py. If past_key_values is on the output weight list, the P-Tuning checkpoint is successfully converted.

Build & Run

Compile the project using CMake:

cmake -B build
cmake --build build -j --config Release

Now you may chat with the quantized ChatGLM-6B model by running:

./build/bin/main -m models/chatglm-ggml.bin -p 你好
# 你好👋！我是人工智能助手 ChatGLM-6B，很高兴见到你，欢迎问我任何问题。

To run the model in interactive mode, add the -i flag. For example:

./build/bin/main -m models/chatglm-ggml.bin -i

In interactive mode, your chat history will serve as the context for the next-round conversation.

Run ./build/bin/main -h to explore more options!

Try Other Models

ChatGLM2-6B

python3 chatglm_cpp/convert.py -i THUDM/chatglm2-6b -t q4_0 -o models/chatglm2-ggml.bin
./build/bin/main -m models/chatglm2-ggml.bin -p 你好 --top_p 0.8 --temp 0.8
# 你好👋！我是人工智能助手 ChatGLM2-6B，很高兴见到你，欢迎问我任何问题。

ChatGLM3-6B

ChatGLM3-6B further supports function call and code interpreter in addition to chat mode.

Chat mode:

python3 chatglm_cpp/convert.py -i THUDM/chatglm3-6b -t q4_0 -o models/chatglm3-ggml.bin
./build/bin/main -m models/chatglm3-ggml.bin -p 你好 --top_p 0.8 --temp 0.8
# 你好👋！我是人工智能助手 ChatGLM3-6B，很高兴见到你，欢迎问我任何问题。

Setting system prompt:

./build/bin/main -m models/chatglm3-ggml.bin -p 你好 -s "You are ChatGLM3, a large language model trained by Zhipu.AI. Follow the user's instructions carefully. Respond using markdown."
# 你好👋！我是 ChatGLM3，有什么问题可以帮您解答吗？

Function call:

$ ./build/bin/main -m models/chatglm3-ggml.bin --top_p 0.8 --temp 0.8 --sp examples/system/function_call.txt -i
System   > Answer the following questions as best as you can. You have access to the following tools: ...
Prompt   > 生成一个随机数
ChatGLM3 > random_number_generator
```python
tool_call(seed=42, range=(0, 100))
```
Tool Call   > Please manually call function `random_number_generator` with args `tool_call(seed=42, range=(0, 100))` and provide the results below.
Observation > 23
ChatGLM3 > 根据您的要求，我使用随机数生成器API生成了一个随机数。根据API返回结果，生成的随机数为23。

Code interpreter:

$ ./build/bin/main -m models/chatglm3-ggml.bin --top_p 0.8 --temp 0.8 --sp examples/system/code_interpreter.txt -i
System   > 你是一位智能AI助手，你叫ChatGLM，你连接着一台电脑，但请注意不能联网。在使用Python解决任务时，你可以运行代码并得到结果，如果运行结果有错误，你需要尽可能对代码进行改进。你可以处理用户上传到电脑上的文件，文件默认存储路径是/mnt/data/。
Prompt   > 列出100以内的所有质数
ChatGLM3 > 好的，我会为您列出100以内的所有质数。
```python
def is_prime(n):
   """Check if a number is prime."""
   if n <= 1:
       return False
   if n <= 3:
       return True
   if n % 2 == 0 or n % 3 == 0:
       return False
   i = 5
   while i * i <= n:
       if n % i == 0 or n % (i + 2) == 0:
           return False
       i += 6
   return True

primes_upto_100 = [i for i in range(2, 101) if is_prime(i)]
primes_upto_100
```

Code Interpreter > Please manually run the code and provide the results below.
Observation      > [2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37, 41, 43, 47, 53, 59, 61, 67, 71, 73, 79, 83, 89, 97]
ChatGLM3 > 100以内的所有质数为：

$$
2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37, 41, 43, 47, 53, 59, 61, 67, 71, 73, 79, 83, 89, 97 
$$

ChatGLM4-9B

Chat mode:

python3 chatglm_cpp/convert.py -i THUDM/glm-4-9b-chat -t q4_0 -o models/chatglm4-ggml.bin
./build/bin/main -m models/chatglm4-ggml.bin -p 你好 --top_p 0.8 --temp 0.8
# 你好👋！有什么可以帮助你的吗？

CodeGeeX2

$ python3 chatglm_cpp/convert.py -i THUDM/codegeex2-6b -t q4_0 -o models/codegeex2-ggml.bin
$ ./build/bin/main -m models/codegeex2-ggml.bin --temp 0 --mode generate -p "\
# language: Python
# write a bubble sort function
"


def bubble_sort(lst):
    for i in range(len(lst) - 1):
        for j in range(len(lst) - 1 - i):
            if lst[j] > lst[j + 1]:
                lst[j], lst[j + 1] = lst[j + 1], lst[j]
    return lst


print(bubble_sort([5, 4, 3, 2, 1]))

Baichuan-13B-Chat

python3 chatglm_cpp/convert.py -i baichuan-inc/Baichuan-13B-Chat -t q4_0 -o models/baichuan-13b-chat-ggml.bin
./build/bin/main -m models/baichuan-13b-chat-ggml.bin -p 你好 --top_k 5 --top_p 0.85 --temp 0.3 --repeat_penalty 1.1
# 你好！有什么我可以帮助你的吗？

Baichuan2-7B-Chat

python3 chatglm_cpp/convert.py -i baichuan-inc/Baichuan2-7B-Chat -t q4_0 -o models/baichuan2-7b-chat-ggml.bin
./build/bin/main -m models/baichuan2-7b-chat-ggml.bin -p 你好 --top_k 5 --top_p 0.85 --temp 0.3 --repeat_penalty 1.05
# 你好！很高兴为您提供帮助。请问有什么问题我可以帮您解答？

Baichuan2-13B-Chat

python3 chatglm_cpp/convert.py -i baichuan-inc/Baichuan2-13B-Chat -t q4_0 -o models/baichuan2-13b-chat-ggml.bin
./build/bin/main -m models/baichuan2-13b-chat-ggml.bin -p 你好 --top_k 5 --top_p 0.85 --temp 0.3 --repeat_penalty 1.05
# 你好！今天我能为您提供什么帮助？

InternLM-Chat-7B

python3 chatglm_cpp/convert.py -i internlm/internlm-chat-7b -t q4_0 -o models/internlm-chat-7b-ggml.bin
./build/bin/main -m models/internlm-chat-7b-ggml.bin -p 你好 --top_p 0.8 --temp 0.8
# 你好，我是书生·浦语，有什么可以帮助你的吗？

InternLM-Chat-20B

python3 chatglm_cpp/convert.py -i internlm/internlm-chat-20b -t q4_0 -o models/internlm-chat-20b-ggml.bin
./build/bin/main -m models/internlm-chat-20b-ggml.bin -p 你好 --top_p 0.8 --temp 0.8
# 你好！有什么我可以帮到你的吗？

Using BLAS

BLAS library can be integrated to further accelerate matrix multiplication. However, in some cases, using BLAS may cause performance degradation. Whether to turn on BLAS should depend on the benchmarking result.

Accelerate Framework

Accelerate Framework is automatically enabled on macOS. To disable it, add the CMake flag -DGGML_NO_ACCELERATE=ON.

OpenBLAS

OpenBLAS provides acceleration on CPU. Add the CMake flag -DGGML_OPENBLAS=ON to enable it.

cmake -B build -DGGML_OPENBLAS=ON && cmake --build build -j

cuBLAS

cuBLAS uses NVIDIA GPU to accelerate BLAS. Add the CMake flag -DGGML_CUBLAS=ON to enable it.

cmake -B build -DGGML_CUBLAS=ON && cmake --build build -j

By default, all kernels will be compiled for all possible CUDA architectures and it takes some time. To run on a specific type of device, you may specify CUDA_ARCHITECTURES to speed up the nvcc compilation. For example:

cmake -B build -DGGML_CUBLAS=ON -DCUDA_ARCHITECTURES="80"       # for A100
cmake -B build -DGGML_CUBLAS=ON -DCUDA_ARCHITECTURES="70;75"    # compatible with both V100 and T4

To find out the CUDA architecture of your GPU device, see Your GPU Compute Capability.

Metal

MPS (Metal Performance Shaders) allows computation to run on powerful Apple Silicon GPU. Add the CMake flag -DGGML_METAL=ON to enable it.

cmake -B build -DGGML_METAL=ON && cmake --build build -j

Python Binding

The Python binding provides high-level chat and stream_chat interface similar to the original Hugging Face ChatGLM(2)-6B.

Installation

Install from PyPI (recommended): will trigger compilation on your platform.

pip install -U chatglm-cpp

To enable cuBLAS acceleration on NVIDIA GPU:

CMAKE_ARGS="-DGGML_CUBLAS=ON" pip install -U chatglm-cpp

To enable Metal on Apple silicon devices:

CMAKE_ARGS="-DGGML_METAL=ON" pip install -U chatglm-cpp

You may also install from source. Add the corresponding CMAKE_ARGS for acceleration.

# install from the latest source hosted on GitHub
pip install git+https://github.com/li-plus/chatglm.cpp.git@main
# or install from your local source after git cloning the repo
pip install .

Pre-built wheels for CPU backend on Linux / MacOS / Windows are published on release. For CUDA / Metal backends, please compile from source code or source distribution.

Using Pre-converted GGML Models

Here is a simple demo that uses chatglm_cpp.Pipeline to load the GGML model and chat with it. First enter the examples folder (cd examples) and launch a Python interactive shell:

>>> import chatglm_cpp
>>> 
>>> pipeline = chatglm_cpp.Pipeline("../models/chatglm-ggml.bin")
>>> pipeline.chat([chatglm_cpp.ChatMessage(role="user", content="你好")])
ChatMessage(role="assistant", content="你好👋！我是人工智能助手 ChatGLM-6B，很高兴见到你，欢迎问我任何问题。", tool_calls=[])

To chat in stream, run the below Python example:

python3 cli_demo.py -m ../models/chatglm-ggml.bin -i

Launch a web demo to chat in your browser:

python3 web_demo.py -m ../models/chatglm-ggml.bin

For other models:

ChatGLM2-6B

python3 cli_demo.py -m ../models/chatglm2-ggml.bin -p 你好 --temp 0.8 --top_p 0.8  # CLI demo
python3 web_demo.py -m ../models/chatglm2-ggml.bin --temp 0.8 --top_p 0.8  # web demo

ChatGLM3-6B

CLI Demo

Chat mode:

python3 cli_demo.py -m ../models/chatglm3-ggml.bin -p 你好 --temp 0.8 --top_p 0.8

Function call:

python3 cli_demo.py -m ../models/chatglm3-ggml.bin --temp 0.8 --top_p 0.8 --sp system/function_call.txt -i

Code interpreter:

python3 cli_demo.py -m ../models/chatglm3-ggml.bin --temp 0.8 --top_p 0.8 --sp system/code_interpreter.txt -i

Web Demo

Install Python dependencies and the IPython kernel for code interpreter.

pip install streamlit jupyter_client ipython ipykernel
ipython kernel install --name chatglm3-demo --user

Launch the web demo:

streamlit run chatglm3_demo.py

Function Call	Code Interpreter

ChatGLM4-9B

Chat mode:

python3 cli_demo.py -m ../models/chatglm4-ggml.bin -p 你好 --temp 0.8 --top_p 0.8

CodeGeeX2

# CLI demo
python3 cli_demo.py -m ../models/codegeex2-ggml.bin --temp 0 --mode generate -p "\
# language: Python
# write a bubble sort function
"
# web demo
python3 web_demo.py -m ../models/codegeex2-ggml.bin --temp 0 --max_length 512 --mode generate --plain

Baichuan-13B-Chat

python3 cli_demo.py -m ../models/baichuan-13b-chat-ggml.bin -p 你好 --top_k 5 --top_p 0.85 --temp 0.3 --repeat_penalty 1.1 # CLI demo
python3 web_demo.py -m ../models/baichuan-13b-chat-ggml.bin --top_k 5 --top_p 0.85 --temp 0.3 --repeat_penalty 1.1   # web demo

Baichuan2-7B-Chat

python3 cli_demo.py -m ../models/baichuan2-7b-chat-ggml.bin -p 你好 --top_k 5 --top_p 0.85 --temp 0.3 --repeat_penalty 1.05 # CLI demo
python3 web_demo.py -m ../models/baichuan2-7b-chat-ggml.bin --top_k 5 --top_p 0.85 --temp 0.3 --repeat_penalty 1.05   # web demo

Baichuan2-13B-Chat

python3 cli_demo.py -m ../models/baichuan2-13b-chat-ggml.bin -p 你好 --top_k 5 --top_p 0.85 --temp 0.3 --repeat_penalty 1.05 # CLI demo
python3 web_demo.py -m ../models/baichuan2-13b-chat-ggml.bin --top_k 5 --top_p 0.85 --temp 0.3 --repeat_penalty 1.05   # web demo

InternLM-Chat-7B

python3 cli_demo.py -m ../models/internlm-chat-7b-ggml.bin -p 你好 --top_p 0.8 --temp 0.8  # CLI demo
python3 web_demo.py -m ../models/internlm-chat-7b-ggml.bin --top_p 0.8 --temp 0.8  # web demo

InternLM-Chat-20B

python3 cli_demo.py -m ../models/internlm-chat-20b-ggml.bin -p 你好 --top_p 0.8 --temp 0.8 # CLI demo
python3 web_demo.py -m ../models/internlm-chat-20b-ggml.bin --top_p 0.8 --temp 0.8 # web demo

Converting Hugging Face LLMs at Runtime

Sometimes it might be inconvenient to convert and save the intermediate GGML models beforehand. Here is an option to directly load from the original Hugging Face model, quantize it into GGML models in a minute, and start serving. All you need is to replace the GGML model path with the Hugging Face model name or path.

>>> import chatglm_cpp
>>> 
>>> pipeline = chatglm_cpp.Pipeline("THUDM/chatglm-6b", dtype="q4_0")
Loading checkpoint shards: 100%|██████████████████████████████████| 8/8 [00:10<00:00,  1.27s/it]
Processing model states: 100%|████████████████████████████████| 339/339 [00:23<00:00, 14.73it/s]
...
>>> pipeline.chat([chatglm_cpp.ChatMessage(role="user", content="你好")])
ChatMessage(role="assistant", content="你好👋！我是人工智能助手 ChatGLM-6B，很高兴见到你，欢迎问我任何问题。", tool_calls=[])

Likewise, replace the GGML model path with Hugging Face model in any example script, and it just works. For example:

python3 cli_demo.py -m THUDM/chatglm-6b -p 你好 -i

API Server

We support various kinds of API servers to integrate with popular frontends. Extra dependencies can be installed by:

pip install 'chatglm-cpp[api]'

Remember to add the corresponding CMAKE_ARGS to enable acceleration.

LangChain API

Start the api server for LangChain:

MODEL=./models/chatglm2-ggml.bin uvicorn chatglm_cpp.langchain_api:app --host 127.0.0.1 --port 8000

Test the api endpoint with curl:

curl http://127.0.0.1:8000 -H 'Content-Type: application/json' -d '{"prompt": "你好"}'

Run with LangChain:

>>> from langchain.llms import ChatGLM
>>> 
>>> llm = ChatGLM(endpoint_url="http://127.0.0.1:8000")
>>> llm.predict("你好")
'你好👋！我是人工智能助手 ChatGLM2-6B，很高兴见到你，欢迎问我任何问题。'

For more options, please refer to examples/langchain_client.py and LangChain ChatGLM Integration.

OpenAI API

Start an API server compatible with OpenAI chat completions protocol:

MODEL=./models/chatglm3-ggml.bin uvicorn chatglm_cpp.openai_api:app --host 127.0.0.1 --port 8000

Test your endpoint with curl:

curl http://127.0.0.1:8000/v1/chat/completions -H 'Content-Type: application/json' \
    -d '{"messages": [{"role": "user", "content": "你好"}]}'

Use the OpenAI client to chat with your model:

>>> from openai import OpenAI
>>> 
>>> client = OpenAI(base_url="http://127.0.0.1:8000/v1")
>>> response = client.chat.completions.create(model="default-model", messages=[{"role": "user", "content": "你好"}])
>>> response.choices[0].message.content
'你好👋！我是人工智能助手 ChatGLM3-6B，很高兴见到你，欢迎问我任何问题。'

For stream response, check out the example client script:

OPENAI_BASE_URL=http://127.0.0.1:8000/v1 python3 examples/openai_client.py --stream --prompt 你好

Tool calling is also supported:

OPENAI_BASE_URL=http://127.0.0.1:8000/v1 python3 examples/openai_client.py --tool_call --prompt 上海天气怎么样

With this API server as backend, ChatGLM.cpp models can be seamlessly integrated into any frontend that uses OpenAI-style API, including mckaywrigley/chatbot-ui, fuergaosi233/wechat-chatgpt, Yidadaa/ChatGPT-Next-Web, and more.

Using Docker

Option 1: Building Locally

Building docker image locally and start a container to run inference on CPU:

docker build . --network=host -t chatglm.cpp
# cpp demo
docker run -it --rm -v $PWD/models:/chatglm.cpp/models chatglm.cpp ./build/bin/main -m models/chatglm-ggml.bin -p "你好"
# python demo
docker run -it --rm -v $PWD/models:/chatglm.cpp/models chatglm.cpp python3 examples/cli_demo.py -m models/chatglm-ggml.bin -p "你好"
# langchain api server
docker run -it --rm -v $PWD/models:/chatglm.cpp/models -p 8000:8000 -e MODEL=models/chatglm-ggml.bin chatglm.cpp \
    uvicorn chatglm_cpp.langchain_api:app --host 0.0.0.0 --port 8000
# openai api server
docker run -it --rm -v $PWD/models:/chatglm.cpp/models -p 8000:8000 -e MODEL=models/chatglm-ggml.bin chatglm.cpp \
    uvicorn chatglm_cpp.openai_api:app --host 0.0.0.0 --port 8000

For CUDA support, make sure nvidia-docker is installed. Then run:

docker build . --network=host -t chatglm.cpp-cuda \
    --build-arg BASE_IMAGE=nvidia/cuda:12.2.0-devel-ubuntu20.04 \
    --build-arg CMAKE_ARGS="-DGGML_CUBLAS=ON -DCUDA_ARCHITECTURES=80"
docker run -it --rm --gpus all -v $PWD/models:/chatglm.cpp/models chatglm.cpp-cuda \
    ./build/bin/main -m models/chatglm-ggml.bin -p "你好"

Option 2: Using Pre-built Image

The pre-built image for CPU inference is published on both Docker Hub and GitHub Container Registry (GHCR).

To pull from Docker Hub and run demo:

docker run -it --rm -v $PWD/models:/chatglm.cpp/models liplusx/chatglm.cpp:main \
    ./build/bin/main -m models/chatglm-ggml.bin -p "你好"

To pull from GHCR and run demo:

docker run -it --rm -v $PWD/models:/chatglm.cpp/models ghcr.io/li-plus/chatglm.cpp:main \
    ./build/bin/main -m models/chatglm-ggml.bin -p "你好"

Python demo and API servers are also supported in pre-built image. Use it in the same way as Option 1.

Performance

Environment:

CPU backend performance is measured on a Linux server with Intel(R) Xeon(R) Platinum 8260 CPU @ 2.40GHz using 16 threads.
CUDA backend is measured on a V100-SXM2-32GB GPU using 1 thread.
MPS backend is measured on an Apple M2 Ultra device using 1 thread.

ChatGLM-6B:

	Q4_0	Q4_1	Q5_0	Q5_1	Q8_0	F16
ms/token (CPU @ Platinum 8260)	74	77	86	89	114	189
ms/token (CUDA @ V100 SXM2)	8.1	8.7	9.4	9.5	12.0	19.1
ms/token (MPS @ M2 Ultra)	11.5	12.3	N/A	N/A	16.1	24.4
file size	3.3G	3.7G	4.0G	4.4G	6.2G	12G
mem usage	4.0G	4.4G	4.7G	5.1G	6.9G	13G

ChatGLM2-6B / ChatGLM3-6B / CodeGeeX2:

	Q4_0	Q4_1	Q5_0	Q5_1	Q8_0	F16
ms/token (CPU @ Platinum 8260)	64	71	79	83	106	189
ms/token (CUDA @ V100 SXM2)	7.9	8.3	9.2	9.2	11.7	18.5
ms/token (MPS @ M2 Ultra)	10.0	10.8	N/A	N/A	14.5	22.2
file size	3.3G	3.7G	4.0G	4.4G	6.2G	12G
mem usage	3.4G	3.8G	4.1G	4.5G	6.2G	12G

ChatGLM4-9B:

	Q4_0	Q4_1	Q5_0	Q5_1	Q8_0	F16
ms/token (CPU @ Platinum 8260)	105	105	122	134	158	279
ms/token (CUDA @ V100 SXM2)	12.1	12.5	13.8	13.9	17.7	27.7
file size	5.0G	5.5G	6.1G	6.6G	9.4G	18G

Baichuan-7B / Baichuan2-7B:

	Q4_0	Q4_1	Q5_0	Q5_1	Q8_0	F16
ms/token (CPU @ Platinum 8260)	85.3	94.8	103.4	109.6	136.8	248.5
ms/token (CUDA @ V100 SXM2)	8.7	9.2	10.2	10.3	13.2	21.0
ms/token (MPS @ M2 Ultra)	11.3	12.0	N/A	N/A	16.4	25.6
file size	4.0G	4.4G	4.9G	5.3G	7.5G	14G
mem usage	4.5G	4.9G	5.3G	5.7G	7.8G	14G

Baichuan-13B / Baichuan2-13B:

	Q4_0	Q4_1	Q5_0	Q5_1	Q8_0	F16
ms/token (CPU @ Platinum 8260)	161.7	175.8	189.9	192.3	255.6	459.6
ms/token (CUDA @ V100 SXM2)	13.7	15.1	16.3	16.9	21.9	36.8
ms/token (MPS @ M2 Ultra)	18.2	18.8	N/A	N/A	27.2	44.4
file size	7.0G	7.8G	8.5G	9.3G	14G	25G
mem usage	7.8G	8.8G	9.5G	10G	14G	25G

InternLM-7B:

	Q4_0	Q4_1	Q5_0	Q5_1	Q8_0	F16
ms/token (CPU @ Platinum 8260)	85.3	90.1	103.5	112.5	137.3	232.2
ms/token (CUDA @ V100 SXM2)	9.1	9.4	10.5	10.5	13.3	21.1

InternLM-20B:

	Q4_0	Q4_1	Q5_0	Q5_1	Q8_0	F16
ms/token (CPU @ Platinum 8260)	230.0	236.7	276.6	290.6	357.1	N/A
ms/token (CUDA @ V100 SXM2)	21.6	23.2	25.0	25.9	33.4	N/A

Model Quality

We measure model quality by evaluating the perplexity over the WikiText-2 test dataset, following the strided sliding window strategy in https://huggingface.co/docs/transformers/perplexity. Lower perplexity usually indicates a better model.

Download and unzip the dataset from link. Measure the perplexity with a stride of 512 and max input length of 2048:

./build/bin/perplexity -m models/chatglm3-base-ggml.bin -f wikitext-2-raw/wiki.test.raw -s 512 -l 2048

	Q4_0	Q4_1	Q5_0	Q5_1	Q8_0	F16
ChatGLM3-6B-Base	6.215	6.184	5.997	6.015	5.965	5.971
ChatGLM4-9B-Base	6.851	6.793	6.652	6.635	6.582	6.586

Development

Unit Test & Benchmark

To perform unit tests, add this CMake flag -DCHATGLM_ENABLE_TESTING=ON to enable testing. Recompile and run the unit test (including benchmark).

mkdir -p build && cd build
cmake .. -DCHATGLM_ENABLE_TESTING=ON && make -j
./bin/chatglm_test

For benchmark only:

./bin/chatglm_test --gtest_filter='Benchmark.*'

Lint

To format the code, run make lint inside the build folder. You should have clang-format, black and isort pre-installed.

Performance

To detect the performance bottleneck, add the CMake flag -DGGML_PERF=ON:

cmake .. -DGGML_PERF=ON && make -j

This will print timing for each graph operation when running the model.

Acknowledgements

This project is greatly inspired by @ggerganov's llama.cpp and is based on his NN library ggml.
Thank @THUDM for the amazing ChatGLM-6B, ChatGLM2-6B, ChatGLM3 and GLM-4 and for releasing the model sources and checkpoints.

everydoc/chatglm.cpp

ChatGLM.cpp

Features

Getting Started

Using BLAS

Python Binding

API Server

Using Docker

Performance

Model Quality

Development

Acknowledgements