|
--- |
|
language: |
|
- zh |
|
- en |
|
tags: |
|
- glm |
|
- chatglm |
|
- ggml |
|
--- |
|
# ChatGLM3-6B-int4 |
|
|
|
介绍 (Introduction) |
|
ChatGLM3-6B 是 ChatGLM 系列最新一代的开源模型，[THUDM/chatglm3-6b](https://huggingface.co/THUDM/chatglm3-6b) |
|
|
|
用 [ChatGLM.CPP]() 基於 GGML quantize 生成 Q4_0、Q4_1 權重 weights 儲存於此倉庫。 |
|
|
|
## Performance |
|
|Model |GGML quantize method| HDD size |1 token\*| |
|
|----------------------|--------------------|----------|---------| |
|
|chatglm3-ggml-q4_0.bin| q4_0 | 3.51 GB | 74ms | |
|
|chatglm3-ggml-q4_1.bin| q4_1 | 3.9 GB | 77ms | |
|
\* ms/token (CPU @ Platinum 8260) from [reference](https://github.com/li-plus/chatglm.cpp#performance) |
|
|
|
## Getting Started |
|
1. Install dependency |
|
```sh |
|
pip install chatglm-cpp transformers |
|
``` |
|
|
|
2. Download weight |
|
```sh |
|
wget https://huggingface.co/npc0/chatglm3-6b-int4/resolve/main/chatglm3-ggml-q4_0.bin |
|
``` |
|
|
|
3. Code |
|
```py |
|
import chatglm_cpp |
|
|
|
pipeline = chatglm_cpp.Pipeline("./chatglm3-ggml-q4_0.bin") |
|
pipeline.chat([chatglm_cpp.ChatMessage(role="user", content="你好")]) |
|
# Output: 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=[]) |
|
``` |
