mirror of
https://github.com/datawhalechina/self-llm.git
synced 2026-08-28 15:11:50 +08:00
docs: add MiniCPM5 tutorials
This commit is contained in:
KMnO4-zx
@@ -85,6 +85,7 @@
|
||||
<table align="center">
|
||||
<tr>
|
||||
<td valign="top" width="25%">
|
||||
• <a href="./support_model.md#minicpm5">MiniCPM5</a><br>
|
||||
• <a href="./support_model.md#kimi-k25">Kimi-K2.5</a><br>
|
||||
• <a href="./support_model.md#step-35-flash">Step-3.5-Flash</a><br>
|
||||
• <a href="./support_model.md#glm-47-flash">GLM-4.7-Flash</a><br>
|
||||
@@ -96,14 +97,14 @@
|
||||
• <a href="./support_model.md#qwen3-vl-4b-instruct">Qwen3-VL</a><br>
|
||||
• <a href="./support_model.md#spatiallm">SpatialLM</a><br>
|
||||
• <a href="./support_model.md#hunyuan3d-2">Hunyuan3D-2</a><br>
|
||||
• <a href="./support_model.md#qwen2-vl">Qwen2-VL</a><br>
|
||||
• <a href="./support_model.md#qwen2-vl">Qwen2-VL</a>
|
||||
</td>
|
||||
<td valign="top" width="25%">
|
||||
• <a href="./support_model.md#minicpm-o-2_6">MiniCPM-o</a><br>
|
||||
• <a href="./support_model.md#qwen25-coder">Qwen2.5-Coder</a><br>
|
||||
• <a href="./support_model.md#deepseek-coder-v2">DeepSeek-Coder-V2</a><br>
|
||||
• <a href="./support_model.md#gpt-oss-20b">gpt-oss-20b</a><br>
|
||||
• <a href="./support_model.md#glm-41-thinking">GLM-4.1-Thinking</a>
|
||||
</td>
|
||||
<td valign="top" width="25%">
|
||||
• <a href="./support_model.md#glm-41-thinking">GLM-4.1-Thinking</a><br>
|
||||
• <a href="./support_model.md#deepseek-r1-distill">DeepSeek-R1</a><br>
|
||||
• <a href="./support_model.md#internlm3">InternLM3</a><br>
|
||||
• <a href="./support_model.md#phi4">phi4</a><br>
|
||||
@@ -112,12 +113,12 @@
|
||||
• <a href="./support_model.md#deepseek-深度求索">DeepSeek</a><br>
|
||||
• <a href="./support_model.md#baichuan-百川智能">Baichuan</a><br>
|
||||
• <a href="./support_model.md#internlm">InternLM</a><br>
|
||||
• <a href="./support_model.md#kimi">Kimi</a><br>
|
||||
• <a href="./support_model.md#ernie-45">ERNIE-4.5</a><br>
|
||||
• <a href="./support_model.md#llama4">Llama4</a><br>
|
||||
• <a href="./support_model.md#apple-openelm">Apple OpenELM</a>
|
||||
• <a href="./support_model.md#kimi">Kimi</a>
|
||||
</td>
|
||||
<td valign="top" width="25%">
|
||||
• <a href="./support_model.md#ernie-45">ERNIE-4.5</a><br>
|
||||
• <a href="./support_model.md#llama4">Llama4</a><br>
|
||||
• <a href="./support_model.md#apple-openelm">Apple OpenELM</a><br>
|
||||
• <a href="./support_model.md#llama31-8b-instruct">Llama3.1</a><br>
|
||||
• <a href="./support_model.md#gemma-2-9b-it">Gemma-2</a><br>
|
||||
• <a href="./support_model.md#qwen25">Qwen2.5</a><br>
|
||||
@@ -128,10 +129,10 @@
|
||||
• <a href="./support_model.md#minicpm">MiniCPM</a><br>
|
||||
• <a href="./support_model.md#yi-零一万物">Yi 零一万物</a><br>
|
||||
• <a href="./support_model.md#yuan20">Yuan2.0</a><br>
|
||||
• <a href="./support_model.md#yuan20-m32">Yuan2.0-M32</a><br>
|
||||
• <a href="./support_model.md#哔哩哔哩-index-19b">哔哩哔哩 Index</a>
|
||||
• <a href="./support_model.md#yuan20-m32">Yuan2.0-M32</a>
|
||||
</td>
|
||||
<td valign="top" width="25%">
|
||||
• <a href="./support_model.md#哔哩哔哩-index-19b">哔哩哔哩 Index</a><br>
|
||||
• <a href="./support_model.md#characterglm-6b">CharacterGLM</a><br>
|
||||
• <a href="./support_model.md#bluelm-vivo-蓝心大模型">BlueLM</a><br>
|
||||
• <a href="./support_model.md#qwen-audio">Qwen-Audio</a><br>
|
||||
|
||||
+16
-11
@@ -85,21 +85,26 @@
|
||||
<table align="center">
|
||||
<tr>
|
||||
<td valign="top" width="25%">
|
||||
• <a href="./support_model.md#minicpm5">MiniCPM5</a><br>
|
||||
• <a href="./support_model.md#kimi-k25">Kimi-K2.5</a><br>
|
||||
• <a href="./support_model.md#step-35-flash">Step-3.5-Flash</a><br>
|
||||
• <a href="./support_model.md#glm-47-flash">GLM-4.7-Flash</a><br>
|
||||
• <a href="./support_model.md#gemma3">Gemma3</a><br>
|
||||
• <a href="./support_model.md#minimax-m3">MiniMax-M3</a><br>
|
||||
• <a href="./support_model.md#minimax-m25">MiniMax-M2.5</a><br>
|
||||
• <a href="./support_model.md#minimax-m2">MiniMax-M2</a><br>
|
||||
• <a href="./support_model.md#qwen3">Qwen3</a><br>
|
||||
• <a href="./support_model.md#qwen3-vl-4b-instruct">Qwen3-VL</a><br>
|
||||
• <a href="./support_model.md#spatiallm">SpatialLM</a><br>
|
||||
• <a href="./support_model.md#hunyuan3d-2">Hunyuan3D-2</a><br>
|
||||
• <a href="./support_model.md#qwen2-vl">Qwen2-VL</a><br>
|
||||
• <a href="./support_model.md#qwen2-vl">Qwen2-VL</a>
|
||||
</td>
|
||||
<td valign="top" width="25%">
|
||||
• <a href="./support_model.md#minicpm-o-2_6">MiniCPM-o</a><br>
|
||||
• <a href="./support_model.md#qwen25-coder">Qwen2.5-Coder</a><br>
|
||||
• <a href="./support_model.md#deepseek-coder-v2">DeepSeek-Coder-V2</a><br>
|
||||
• <a href="./support_model.md#gpt-oss-20b">gpt-oss-20b</a><br>
|
||||
• <a href="./support_model.md#glm-41-thinking">GLM-4.1-Thinking</a>
|
||||
</td>
|
||||
<td valign="top" width="25%">
|
||||
• <a href="./support_model.md#glm-41-thinking">GLM-4.1-Thinking</a><br>
|
||||
• <a href="./support_model.md#deepseek-r1-distill">DeepSeek-R1</a><br>
|
||||
• <a href="./support_model.md#internlm3">InternLM3</a><br>
|
||||
• <a href="./support_model.md#phi4">phi4</a><br>
|
||||
@@ -108,12 +113,12 @@
|
||||
• <a href="./support_model.md#deepseek-深度求索">DeepSeek</a><br>
|
||||
• <a href="./support_model.md#baichuan-百川智能">Baichuan</a><br>
|
||||
• <a href="./support_model.md#internlm">InternLM</a><br>
|
||||
• <a href="./support_model.md#kimi">Kimi</a><br>
|
||||
• <a href="./support_model.md#ernie-45">ERNIE-4.5</a><br>
|
||||
• <a href="./support_model.md#llama4">Llama4</a><br>
|
||||
• <a href="./support_model.md#apple-openelm">Apple OpenELM</a>
|
||||
• <a href="./support_model.md#kimi">Kimi</a>
|
||||
</td>
|
||||
<td valign="top" width="25%">
|
||||
• <a href="./support_model.md#ernie-45">ERNIE-4.5</a><br>
|
||||
• <a href="./support_model.md#llama4">Llama4</a><br>
|
||||
• <a href="./support_model.md#apple-openelm">Apple OpenELM</a><br>
|
||||
• <a href="./support_model.md#llama31-8b-instruct">Llama3.1</a><br>
|
||||
• <a href="./support_model.md#gemma-2-9b-it">Gemma-2</a><br>
|
||||
• <a href="./support_model.md#qwen25">Qwen2.5</a><br>
|
||||
@@ -124,10 +129,10 @@
|
||||
• <a href="./support_model.md#minicpm">MiniCPM</a><br>
|
||||
• <a href="./support_model.md#yi-零一万物">Yi 零一万物</a><br>
|
||||
• <a href="./support_model.md#yuan20">Yuan2.0</a><br>
|
||||
• <a href="./support_model.md#yuan20-m32">Yuan2.0-M32</a><br>
|
||||
• <a href="./support_model.md#哔哩哔哩-index-19b">Bilibili Index</a>
|
||||
• <a href="./support_model.md#yuan20-m32">Yuan2.0-M32</a>
|
||||
</td>
|
||||
<td valign="top" width="25%">
|
||||
• <a href="./support_model.md#哔哩哔哩-index-19b">Bilibili Index</a><br>
|
||||
• <a href="./support_model.md#characterglm-6b">CharacterGLM</a><br>
|
||||
• <a href="./support_model.md#bluelm-vivo-蓝心大模型">BlueLM</a><br>
|
||||
• <a href="./support_model.md#qwen-audio">Qwen-Audio</a><br>
|
||||
@@ -293,4 +298,4 @@
|
||||
|
||||
<div align=center style="margin-top: 30px;">
|
||||
<img src="./images/star-history-20251220.png"/>
|
||||
</div>
|
||||
</div>
|
||||
|
||||
@@ -0,0 +1,227 @@
|
||||
# 01-MiniCPM5-1B vLLM 部署调用
|
||||
|
||||
## vLLM 简介
|
||||
|
||||
`vLLM` 框架是一个高效的大语言模型**推理和部署服务系统**，具备以下特性：
|
||||
|
||||
- **高效的内存管理**：通过 `PagedAttention` 算法，`vLLM` 实现了对 `KV` 缓存的高效管理，减少了内存浪费，优化了模型的运行效率。
|
||||
- **高吞吐量**：`vLLM` 支持异步处理和连续批处理请求，显著提高了模型推理的吞吐量。
|
||||
- **易用性**：`vLLM` 与 `HuggingFace` 模型无缝集成，兼容 `OpenAI` 的 `API` 服务器。
|
||||
- **开源共享**：`vLLM` 开源，社区活跃。
|
||||
|
||||
> `MiniCPM5-1B` 采用**标准 `LlamaForCausalLM` 架构**，主流推理引擎可直接加载——无需自定义算子、无需模型代码 fork。本教程使用 `vLLM` 部署，**文中启动日志与接口返回均为实测真实输出**。
|
||||
|
||||
## 关于 MiniCPM5-1B
|
||||
|
||||
`MiniCPM5-1B` 是面壁智能（ModelBest）/ OpenBMB 发布的 1B 稠密 Transformer，面向端侧、本地部署与资源受限场景，具备：
|
||||
|
||||
- **同尺寸开源 SOTA**：在 Agentic 工具调用、代码生成、高难推理上优势明显。
|
||||
- **双模式推理（Hybrid Reasoning）**：内置 `<think>` chat template，可通过 `enable_thinking` 在「思考」与「非思考」模式间切换，同一份权重既是快速助手也是深度推理器。
|
||||
- **原生长上下文**：支持最长 128K 上下文。
|
||||
- **架构**：`LlamaForCausalLM`，24 层，hidden_size 1536，GQA（16 注意力头 / 2 KV 头），rope_theta=5000000。
|
||||
|
||||
## 环境准备
|
||||
|
||||
本文实测基础环境如下：
|
||||
|
||||
```
|
||||
----------------
|
||||
ubuntu 22.04
|
||||
python 3.12
|
||||
NVIDIA 驱动 580.105.08
|
||||
GPU: RTX 4090 D (24G)
|
||||
torch 2.11.0+cu128
|
||||
vllm 0.23.0
|
||||
----------------
|
||||
```
|
||||
|
||||
> 本文默认学习者已配置好 `Pytorch (cuda)` 环境，如未配置请先自行安装。
|
||||
|
||||
```bash
|
||||
python -m pip install --upgrade pip
|
||||
pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
|
||||
|
||||
pip install modelscope
|
||||
pip install "transformers>=5.6"
|
||||
pip install "vllm>=0.21"
|
||||
pip install openai
|
||||
```
|
||||
|
||||
> 若启动时报 `ModuleNotFoundError: No module named 'flash_attn.ops'`，通常是环境里装了 `flash-attn-4`（会留下一个空的 `flash_attn` 命名空间包），而 vLLM 的 rotary 模块检测到 `flash_attn` 后会尝试导入其 `.ops` 子模块。解决：`pip uninstall flash-attn-4`，并删除残留的空目录 `rm -rf $(python -c "import site;print(site.getsitepackages()[0])")/flash_attn`，vLLM 会自动回退到自带实现。
|
||||
|
||||
## 模型下载
|
||||
|
||||
使用 modelscope 中的 `snapshot_download` 函数下载模型。
|
||||
|
||||
新建 `model_download.py`：
|
||||
|
||||
```python
|
||||
# model_download.py
|
||||
from modelscope import snapshot_download
|
||||
|
||||
model_dir = snapshot_download('OpenBMB/MiniCPM5-1B', cache_dir='/root/autodl-tmp')
|
||||
print(f"模型下载完成，保存路径为：{model_dir}")
|
||||
```
|
||||
|
||||
然后执行 `python model_download.py`。
|
||||
|
||||
> 注意：记得修改 `cache_dir` 为你的模型下载路径哦~
|
||||
|
||||
## 创建兼容 OpenAI API 接口的服务器
|
||||
|
||||
`MiniCPM5-1B` 兼容 `OpenAI API` 协议。常用启动参数：
|
||||
|
||||
- `--host` / `--port`：地址与端口
|
||||
- `--model`：模型路径
|
||||
- `--served-model-name`：服务对外的模型名称
|
||||
- `--max-model-len`：最大上下文长度（1B 模型在 24G 显存上可设 `4096` 或更大）
|
||||
- `--gpu-memory-utilization`：显存占用比例（1B 模型很小，0.6 即可）
|
||||
- `--trust-remote-code`：信任远程代码
|
||||
|
||||
```bash
|
||||
vllm serve /root/autodl-tmp/OpenBMB/MiniCPM5-1B \
|
||||
--served-model-name MiniCPM5-1B \
|
||||
--max-model-len 4096 \
|
||||
--gpu-memory-utilization 0.6 \
|
||||
--trust-remote-code \
|
||||
--host 0.0.0.0 --port 8000
|
||||
```
|
||||
|
||||
实测启动日志如下（vLLM 识别为 `LlamaForCausalLM`，1B 权重加载仅 0.52s）：
|
||||
|
||||
|
||||
|
||||
```bash
|
||||
(APIServer) INFO [model.py:611] Resolved architecture: LlamaForCausalLM
|
||||
(EngineCore) INFO [core.py:113] Initializing a V1 LLM engine (v0.23.0) ...
|
||||
(EngineCore) INFO [default_loader.py:397] Loading weights took 0.52 seconds
|
||||
(EngineCore) INFO [model_runner.py:319] Model loading took 2.09 GiB and 2.14 seconds
|
||||
(EngineCore) INFO [gpu_worker.py:480] Available KV cache memory: 11.54 GiB
|
||||
(EngineCore) INFO [kv_cache_utils.py:1744] GPU KV cache size: 504,192 tokens
|
||||
(EngineCore) INFO [core.py:306] init engine (profile, create kv cache, warmup model) took 39.05 s (compilation: 18.97 s)
|
||||
(APIServer) INFO: Application startup complete.
|
||||
```
|
||||
|
||||
> 首次启动会触发 `torch.compile` 编译（约 19s），编译结果会缓存，后续启动更快。出现 `Application startup complete.` 即说明服务成功启动。
|
||||
|
||||
- 查看 `curl http://localhost:8000/v1/models`：
|
||||
|
||||
```json
|
||||
{
|
||||
"object": "list",
|
||||
"data": [
|
||||
{
|
||||
"id": "MiniCPM5-1B",
|
||||
"object": "model",
|
||||
"owned_by": "vllm",
|
||||
"root": "/root/autodl-tmp/OpenBMB/MiniCPM5-1B",
|
||||
"max_model_len": 4096
|
||||
}
|
||||
]
|
||||
}
|
||||
```
|
||||
|
||||
### 思考模式与非思考模式
|
||||
|
||||
`MiniCPM5-1B` 内置 `<think>` 模板，可通过 `chat_template_kwargs.enable_thinking` 按**请求**级别控制：
|
||||
|
||||
- **思考模式**（`enable_thinking=true`，推荐 `temperature=0.9, top_p=0.95`）：先输出 `<think> ... </think>` 推理过程，再给出答案
|
||||
- **非思考模式**（`enable_thinking=false`，推荐 `temperature=0.7, top_p=0.95`）：不强制思考，直接回答
|
||||
|
||||
| 模式 | 推荐参数 | enable_thinking |
|
||||
| --- | --- | --- |
|
||||
| Think | `temperature=0.9, top_p=0.95` | `True` |
|
||||
| No Think | `temperature=0.7, top_p=0.95` | `False` |
|
||||
|
||||
### 用 curl 测试 Chat Completions（非思考模式）
|
||||
|
||||
```bash
|
||||
curl http://localhost:8000/v1/chat/completions \
|
||||
-H "Content-Type: application/json" \
|
||||
-d '{
|
||||
"model": "MiniCPM5-1B",
|
||||
"messages": [
|
||||
{"role": "user", "content": "你是谁？用一句话介绍自己。"}
|
||||
],
|
||||
"temperature": 0.7,
|
||||
"top_p": 0.95,
|
||||
"max_tokens": 256,
|
||||
"extra_body": {"chat_template_kwargs": {"enable_thinking": false}}
|
||||
}'
|
||||
```
|
||||
|
||||
实测返回值如下（`content` 中先是简短的 `<think>` 思考，其后是最终回答，`finish_reason` 为 `stop`）：
|
||||
|
||||
```json
|
||||
{
|
||||
"id": "chatcmpl-9c66165de8661ff3",
|
||||
"object": "chat.completion",
|
||||
"model": "MiniCPM5-1B",
|
||||
"choices": [
|
||||
{
|
||||
"index": 0,
|
||||
"message": {
|
||||
"role": "assistant",
|
||||
"content": "<think>\n嗯，用户让我介绍自己，需要一句话说明身份。MiniCPM系列模型是由面壁智能和OpenBMB社区开发的，所以应该直接说明这一点。\n</think>\n\n我是MiniCPM系列模型，由面壁智能（ModelBest）和OpenBMB开源社区开发。"
|
||||
},
|
||||
"finish_reason": "stop"
|
||||
}
|
||||
],
|
||||
"usage": {
|
||||
"prompt_tokens": 15,
|
||||
"completion_tokens": 57,
|
||||
"total_tokens": 72
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
> 实测发现：`MiniCPM5-1B` 即便在非思考模式下，也常在 `content` 开头先输出一段简短的 `<think> ... </think>` 再给出回答（这是该模型后训练形成的习惯）。若需要纯粹的非思考输出，可适当调大 `max_tokens`。
|
||||
|
||||
### 用 Python 脚本请求（思考模式）
|
||||
|
||||
```python
|
||||
# vllm_openai_chat_completions.py
|
||||
from openai import OpenAI
|
||||
|
||||
client = OpenAI(
|
||||
api_key="sk-xxx", # 随便填写，只是为了通过接口参数校验
|
||||
base_url="http://localhost:8000/v1",
|
||||
)
|
||||
|
||||
# 思考模式：模型会先输出推理过程
|
||||
chat_outputs = client.chat.completions.create(
|
||||
model="MiniCPM5-1B",
|
||||
messages=[{"role": "user", "content": "5的阶乘是多少？"}],
|
||||
temperature=0.9,
|
||||
top_p=0.95,
|
||||
extra_body={"chat_template_kwargs": {"enable_thinking": True}},
|
||||
)
|
||||
print(chat_outputs.choices[0].message.content)
|
||||
```
|
||||
|
||||
输出包含 `<think> ... </think>` 思考过程与最终答案：
|
||||
|
||||
```
|
||||
<think>
|
||||
5 的阶乘记作 5!，等于 5 × 4 × 3 × 2 × 1 ...
|
||||
</think>
|
||||
|
||||
5 的阶乘（5!）= 5 × 4 × 3 × 2 × 1 = 120。
|
||||
```
|
||||
|
||||
### 运行时日志
|
||||
|
||||
在请求处理过程中，`API` 后端会持续打印日志与统计信息，便于观测服务状态。实测运行时日志如下：
|
||||
|
||||
|
||||
|
||||
```bash
|
||||
(EngineCore) INFO [core.py:306] init engine (profile, create kv cache, warmup model) took 39.05 s (compilation: 18.97 s)
|
||||
(APIServer) INFO: Application startup complete.
|
||||
(APIServer) INFO: 127.0.0.1:34630 - "POST /v1/chat/completions HTTP/1.1" 200 OK
|
||||
(APIServer) INFO: 127.0.0.1:34660 - "POST /v1/chat/completions HTTP/1.1" 200 OK
|
||||
```
|
||||
|
||||
## 工具调用（Tool Calling）
|
||||
|
||||
`MiniCPM5-1B` 原生支持 XML 风格的工具调用。在 vLLM 中可配合 `--tool-call-parser` 使用（vLLM 较新版本支持 `minicpm5` 解析器），将模型输出的 `<function ... </function>` 转换为 OpenAI 兼容的 `tool_calls`。具体用法可参考 MiniCPM 官方 cookbook。
|
||||
@@ -0,0 +1,255 @@
|
||||
# 02-MiniCPM5-1B SGLang 部署调用
|
||||
|
||||
## SGLang 简介
|
||||
|
||||
`SGLang` 是一款专为大语言模型/多模态模型设计的高性能推理与服务框架：
|
||||
|
||||
- **后端一键启动**：一条命令完成环境适配与服务发布。
|
||||
- **前端无缝对接**：直接沿用 `OpenAI SDK` 或标准 `HTTP` 调用。
|
||||
- **高性能**：支持 `RadixAttention`（前缀复用）、连续批处理、CUDA Graph 等加速技术。
|
||||
|
||||
> `MiniCPM5-1B` 采用**标准 `LlamaForCausalLM` 架构**，SGLang 可直接加载，无需自定义算子。本教程使用 `SGLang` 部署，**文中启动日志与接口返回均为实测真实输出**。
|
||||
|
||||
> 官方提示：工具调用（Tool Calling）场景下，**SGLang 是推荐后端**——MiniCPM5-1B 输出 XML 风格工具调用，SGLang 内置的 `minicpm5` 解析器可将其原生转换为 OpenAI 兼容的 `tool_calls`。
|
||||
|
||||
## 关于 MiniCPM5-1B
|
||||
|
||||
`MiniCPM5-1B` 是面壁智能 / OpenBMB 的 1B 稠密 Transformer，面向端侧与本地部署：标准 `LlamaForCausalLM` 架构（24 层，GQA，128K 上下文），内置 `<think>` 模板支持「思考 / 非思考」双模式（通过 `enable_thinking` 切换）。
|
||||
|
||||
## 环境准备
|
||||
|
||||
本文实测基础环境如下：
|
||||
|
||||
```
|
||||
----------------
|
||||
ubuntu 22.04
|
||||
python 3.12
|
||||
NVIDIA 驱动 580.105.08
|
||||
GPU: RTX 4090 D (24G, sm89)
|
||||
torch 2.11.0+cu128
|
||||
sglang 0.5.13.post1
|
||||
----------------
|
||||
```
|
||||
|
||||
> 本文默认学习者已配置好 `Pytorch (cuda)` 环境，如未配置请先自行安装。
|
||||
|
||||
```bash
|
||||
python -m pip install --upgrade pip
|
||||
pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
|
||||
|
||||
pip install modelscope
|
||||
pip install "transformers>=5.6"
|
||||
pip install openai
|
||||
|
||||
# 安装 sglang（官方建议 sglang[srt]>=0.5.12）
|
||||
pip install "sglang[srt]>=0.5.12"
|
||||
```
|
||||
|
||||
> 若在 RTX 4090（sm89）上启动报 `Could not load any common_ops library! Expected variant: SM89`，说明默认装的 `sglang-kernel` 是 CUDA 13 / sm90+ 构建，需要换成 sm89 兼容版本：
|
||||
> ```bash
|
||||
> pip install sglang-kernel --index-url https://docs.sglang.ai/whl/cu129/
|
||||
> ```
|
||||
|
||||
## 模型下载
|
||||
|
||||
新建 `model_download.py`：
|
||||
|
||||
```python
|
||||
# model_download.py
|
||||
from modelscope import snapshot_download
|
||||
|
||||
model_dir = snapshot_download('OpenBMB/MiniCPM5-1B', cache_dir='/root/autodl-tmp')
|
||||
print(f"模型下载完成，保存路径为：{model_dir}")
|
||||
```
|
||||
|
||||
执行 `python model_download.py`。
|
||||
|
||||
> 注意：记得修改 `cache_dir` 为你的模型下载路径哦~
|
||||
|
||||
## 启动 SGLang 服务
|
||||
|
||||
`MiniCPM5-1B` 为 1B 模型，单张 24G 显卡绰绰有余，无需张量并行。
|
||||
|
||||
### 命令行直接启动
|
||||
|
||||
```bash
|
||||
python3 -m sglang.launch_server \
|
||||
--model-path /root/autodl-tmp/OpenBMB/MiniCPM5-1B \
|
||||
--served-model-name MiniCPM5-1B \
|
||||
--host 0.0.0.0 \
|
||||
--port 8000 \
|
||||
--mem-fraction-static 0.6 \
|
||||
--context-length 4096 \
|
||||
--trust-remote-code
|
||||
```
|
||||
|
||||
> 新版 SGLang 推荐使用 `sglang serve ...` 入口（与 `python -m sglang.launch_server` 等价）。
|
||||
> 若需工具调用，加上 `--tool-call-parser minicpm5`（或 `--tool-call-parser auto`）。
|
||||
|
||||
常用参数：
|
||||
|
||||
- `--model-path`：模型路径
|
||||
- `--served-model-name`：服务对外的模型名称
|
||||
- `--mem-fraction-static`：静态显存占用比例（1B 模型很小，0.6 即可）
|
||||
- `--context-length`：最大上下文长度
|
||||
- `--tp-size`：张量并行数，单卡无需设置
|
||||
- `--trust-remote-code`：信任远程代码
|
||||
|
||||
实测启动日志如下（SGLang 识别为 `LlamaForCausalLM`，权重加载 0.95s）：
|
||||
|
||||
|
||||
|
||||
```bash
|
||||
[22:24:31] Load weight end. elapsed=0.95 s, type=LlamaForCausalLM, avail mem=11.06 GB, mem usage=2.16 GB.
|
||||
[22:24:31] KV Cache is allocated. dtype: torch.bfloat16, #tokens: 251788, K size: 2.88 GB, V size: 2.88 GB
|
||||
[22:24:31] Memory pool end. avail mem=5.18 GB
|
||||
[22:24:31] Capture cuda graph begin. This can take up to several minutes. avail mem=4.73 GB
|
||||
[22:25:18] Capture cuda graph end. Time elapsed: 47.12 s. mem usage=3.97 GB. avail mem=0.76 GB.
|
||||
[22:25:36] INFO: Application startup complete.
|
||||
[22:25:37] The server is fired up and ready to roll!
|
||||
```
|
||||
|
||||
> 首次启动会进行 CUDA graph 捕获（约 47s），完成后出现 `The server is fired up and ready to roll!` 即说明服务成功启动。
|
||||
|
||||
### Python 启动脚本
|
||||
|
||||
```python
|
||||
# start_server.py
|
||||
from sglang.utils import launch_server_cmd, wait_for_server
|
||||
|
||||
cmd = (
|
||||
"python3 -m sglang.launch_server "
|
||||
"--model-path /root/autodl-tmp/OpenBMB/MiniCPM5-1B "
|
||||
"--served-model-name MiniCPM5-1B "
|
||||
"--host 0.0.0.0 --port 8000 "
|
||||
"--mem-fraction-static 0.6 --context-length 4096 "
|
||||
"--trust-remote-code"
|
||||
)
|
||||
|
||||
server_process, port = launch_server_cmd(cmd, port=8000)
|
||||
wait_for_server(f"http://127.0.0.1:{port}")
|
||||
print(f"SGLang Server started: http://127.0.0.1:{port}")
|
||||
```
|
||||
|
||||
## 调用示例
|
||||
|
||||
### 查看模型列表
|
||||
|
||||
```bash
|
||||
curl http://localhost:8000/v1/models
|
||||
```
|
||||
|
||||
实测返回值（`owned_by` 为 `sglang`）：
|
||||
|
||||
```json
|
||||
{
|
||||
"object": "list",
|
||||
"data": [
|
||||
{
|
||||
"id": "MiniCPM5-1B",
|
||||
"object": "model",
|
||||
"owned_by": "sglang",
|
||||
"root": "MiniCPM5-1B",
|
||||
"max_model_len": 4096
|
||||
}
|
||||
]
|
||||
}
|
||||
```
|
||||
|
||||
### 聊天对话（思考模式）
|
||||
|
||||
`MiniCPM5-1B` 内置 `<think>` 模板，通过 `chat_template_kwargs.enable_thinking` 控制模式：
|
||||
|
||||
| 模式 | 推荐参数 | enable_thinking |
|
||||
| --- | --- | --- |
|
||||
| Think | `temperature=0.9, top_p=0.95` | `True` |
|
||||
| No Think | `temperature=0.7, top_p=0.95` | `False` |
|
||||
|
||||
```python
|
||||
# test_chat.py
|
||||
from openai import OpenAI
|
||||
|
||||
client = OpenAI(api_key="EMPTY", base_url="http://localhost:8000/v1")
|
||||
|
||||
# 思考模式：先输出 <think> ... </think>，再给答案
|
||||
response = client.chat.completions.create(
|
||||
model="MiniCPM5-1B",
|
||||
messages=[{"role": "user", "content": "5的阶乘是多少？"}],
|
||||
temperature=0.9,
|
||||
top_p=0.95,
|
||||
max_tokens=768,
|
||||
extra_body={"chat_template_kwargs": {"enable_thinking": True}},
|
||||
)
|
||||
print(response.choices[0].message.content)
|
||||
```
|
||||
|
||||
实测输出包含完整推理与最终答案（`finish_reason: stop`）：
|
||||
|
||||
```
|
||||
<think>
|
||||
5 的阶乘记作 5!，等于 5 × 4 × 3 × 2 × 1 = 120 ...
|
||||
</think>
|
||||
|
||||
5 的阶乘（5!）等于 5 × 4 × 3 × 2 × 1 = 120。
|
||||
```
|
||||
|
||||
### 非思考模式
|
||||
|
||||
```python
|
||||
response = client.chat.completions.create(
|
||||
model="MiniCPM5-1B",
|
||||
messages=[{"role": "user", "content": "你是谁？用一句话介绍自己。"}],
|
||||
temperature=0.7,
|
||||
top_p=0.95,
|
||||
extra_body={"chat_template_kwargs": {"enable_thinking": False}},
|
||||
)
|
||||
print(response.choices[0].message.content)
|
||||
```
|
||||
|
||||
> 实测发现：`MiniCPM5-1B` 即便在非思考模式下，也常在 `content` 开头先输出一段简短的 `<think> ... </think>` 再给出回答，这是该模型后训练形成的习惯。
|
||||
|
||||
### 运行时日志
|
||||
|
||||
请求处理时，SGLang 后端会持续打印解码批次的统计信息。实测运行时日志如下：
|
||||
|
||||
|
||||
|
||||
```bash
|
||||
[22:25:36] INFO: Application startup complete.
|
||||
[22:25:37] The server is fired up and ready to roll!
|
||||
[22:25:37] INFO: 127.0.0.1:xxxxx - "POST /v1/chat/completions HTTP/1.1" 200 OK
|
||||
```
|
||||
|
||||
### 工具调用（Tool Calling）
|
||||
|
||||
SGLang 是 MiniCPM5-1B 工具调用的推荐后端。启动时加 `--tool-call-parser minicpm5`，即可把模型输出的 XML 风格 `<function ... </function>` 原生转换为 OpenAI 兼容的 `tool_calls`：
|
||||
|
||||
```bash
|
||||
python3 -m sglang.launch_server --model-path /root/autodl-tmp/OpenBMB/MiniCPM5-1B \
|
||||
--served-model-name MiniCPM5-1B --port 8000 --tool-call-parser minicpm5
|
||||
```
|
||||
|
||||
```python
|
||||
tools = [{
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "get_weather",
|
||||
"description": "获取指定城市的天气",
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {"city": {"type": "string", "description": "城市名"}},
|
||||
"required": ["city"],
|
||||
},
|
||||
},
|
||||
}]
|
||||
response = client.chat.completions.create(
|
||||
model="MiniCPM5-1B",
|
||||
messages=[{"role": "user", "content": "北京今天天气怎么样？"}],
|
||||
tools=tools,
|
||||
)
|
||||
print(response.choices[0].message.tool_calls)
|
||||
```
|
||||
|
||||
## 小结
|
||||
|
||||
`MiniCPM5-1B` 作为标准 `LlamaForCausalLM` 架构的 1B 模型，在 `vLLM` 与 `SGLang` 中均可一键部署，无需任何特殊算子。结合其「思考/非思考」双模式与原生工具调用能力，非常适合端侧助手、coding agent 与工具调用场景。
|
||||
@@ -0,0 +1,447 @@
|
||||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"id": "mcpm-00",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# MiniCPM5-1B LoRA 微调及 SwanLab 可视化记录\n",
|
||||
"\n",
|
||||
"本教程使用 LoRA 方法在「甄嬛」角色对话数据集上微调 **MiniCPM5-1B**，并使用 **SwanLab** 记录训练过程。\n",
|
||||
"\n",
|
||||
"## 环境配置\n",
|
||||
"\n",
|
||||
"> MiniCPM5-1B 采用标准 `LlamaForCausalLM` 架构，需要 `transformers>=5.6`。\n",
|
||||
"\n",
|
||||
"```bash\n",
|
||||
"pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple\n",
|
||||
"pip install \"transformers>=5.6\" accelerate datasets peft swanlab modelscope\n",
|
||||
"```"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"id": "mcpm-01",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 导入环境"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"id": "mcpm-02",
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import json\n",
|
||||
"import torch\n",
|
||||
"from datasets import Dataset\n",
|
||||
"from transformers import (\n",
|
||||
" AutoTokenizer,\n",
|
||||
" AutoModelForCausalLM,\n",
|
||||
" TrainingArguments,\n",
|
||||
" Trainer,\n",
|
||||
" DataCollatorForSeq2Seq,\n",
|
||||
")\n",
|
||||
"from peft import LoraConfig, TaskType, get_peft_model\n",
|
||||
"import swanlab\n",
|
||||
"from swanlab.integration.transformers import SwanLabCallback\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"id": "mcpm-03",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 读取数据集\n",
|
||||
"\n",
|
||||
"本教程使用甄嬛对话数据集（Alpaca 格式：`instruction / input / output`）。"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"id": "mcpm-04",
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"dataset_path = \"/root/autodl-tmp/huanhuan.json\" # 注意修改为你的数据集路径\n",
|
||||
"with open(dataset_path, \"r\", encoding=\"utf-8\") as f:\n",
|
||||
" data = json.load(f)\n",
|
||||
"\n",
|
||||
"ds = Dataset.from_list(data)\n",
|
||||
"ds"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"id": "mcpm-05",
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"ds[:3]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"id": "mcpm-06",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 认识 MiniCPM5 的 Chat Template\n",
|
||||
"\n",
|
||||
"MiniCPM5-1B 采用 `<|im_start|>role\\n...<|im_end|>\\n` 格式，支持 `enable_thinking` 控制思考模式。角色扮演任务我们**关闭思考模式**。"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"id": "mcpm-07",
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"model_id = \"/root/autodl-tmp/OpenBMB/MiniCPM5-1B\" # 注意修改为你的模型路径\n",
|
||||
"tokenizer = AutoTokenizer.from_pretrained(model_id)\n",
|
||||
"print(\"eos:\", tokenizer.eos_token, tokenizer.eos_token_id)\n",
|
||||
"print(\"pad:\", tokenizer.pad_token, tokenizer.pad_token_id)\n",
|
||||
"print(\"<|im_end|> id:\", tokenizer.convert_tokens_to_ids(\"<|im_end|>\"))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"id": "mcpm-08",
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"messages = [\n",
|
||||
" {\"role\": \"system\", \"content\": \"现在你要扮演皇帝身边的女人--甄嬛\"},\n",
|
||||
" {\"role\": \"user\", \"content\": \"你父亲是谁？\"},\n",
|
||||
" {\"role\": \"assistant\", \"content\": \"家父是大理寺少卿甄远道。\"},\n",
|
||||
"]\n",
|
||||
"\n",
|
||||
"text = tokenizer.apply_chat_template(\n",
|
||||
" messages, tokenize=False, add_generation_prompt=False, enable_thinking=False\n",
|
||||
")\n",
|
||||
"print(text)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"id": "mcpm-09",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 处理数据集\n",
|
||||
"\n",
|
||||
"**注意**：MiniCPM5 模板在「带 generation prompt」时会追加 `<think>\\n\\n</think>\\n\\n`（非思考占位），但「完整对话渲染」时助手回合**不**含这个 think 块。因此**不能用 token 级切片**，需分别对「前缀」和「回答」单独 tokenize 再拼接。`labels` 中只有回答部分参与 loss。"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"id": "mcpm-10",
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def process_func(example):\n",
|
||||
" MAX_LENGTH = 1024\n",
|
||||
" SYS = \"现在你要扮演皇帝身边的女人--甄嬛\"\n",
|
||||
"\n",
|
||||
" messages = [{\"role\": \"system\", \"content\": SYS},\n",
|
||||
" {\"role\": \"user\", \"content\": example[\"instruction\"] + example[\"input\"]}]\n",
|
||||
" # 前缀（system + user，带 generation prompt，含非思考 think 占位），不计算 loss\n",
|
||||
" prompt_ids = tokenizer.apply_chat_template(\n",
|
||||
" messages, tokenize=True, add_generation_prompt=True,\n",
|
||||
" enable_thinking=False, return_dict=False,\n",
|
||||
" )\n",
|
||||
" # 回答部分：output + 结束符 <|im_end|>\n",
|
||||
" response_ids = tokenizer(example[\"output\"], add_special_tokens=False).input_ids \\\n",
|
||||
" + [tokenizer.convert_tokens_to_ids(\"<|im_end|>\")]\n",
|
||||
"\n",
|
||||
" input_ids = prompt_ids + response_ids\n",
|
||||
" labels = [-100] * len(prompt_ids) + response_ids\n",
|
||||
" attention_mask = [1] * len(input_ids)\n",
|
||||
"\n",
|
||||
" if len(input_ids) > MAX_LENGTH: # 超长截断\n",
|
||||
" input_ids = input_ids[:MAX_LENGTH]\n",
|
||||
" attention_mask = attention_mask[:MAX_LENGTH]\n",
|
||||
" labels = labels[:MAX_LENGTH]\n",
|
||||
"\n",
|
||||
" return {\"input_ids\": input_ids, \"attention_mask\": attention_mask, \"labels\": labels}"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"id": "mcpm-11",
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"tokenized_id = ds.map(process_func, remove_columns=ds.column_names)\n",
|
||||
"tokenized_id"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"id": "mcpm-12",
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"# 查看完整输入\n",
|
||||
"print(tokenizer.decode(tokenized_id[0][\"input_ids\"]))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"id": "mcpm-13",
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"# 查看 labels（过滤掉 -100 后即为模型需要学习的回答）\n",
|
||||
"print(tokenizer.decode(list(filter(lambda x: x != -100, tokenized_id[0][\"labels\"]))))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"id": "mcpm-14",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 加载模型\n",
|
||||
"\n",
|
||||
"MiniCPM5-1B 是标准 `LlamaForCausalLM`，直接用 `AutoModelForCausalLM` 加载。"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"id": "mcpm-15",
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"model = AutoModelForCausalLM.from_pretrained(\n",
|
||||
" model_id,\n",
|
||||
" dtype=torch.bfloat16,\n",
|
||||
" device_map=\"auto\",\n",
|
||||
")\n",
|
||||
"# 开启梯度检查点时需要调用该方法\n",
|
||||
"model.enable_input_require_grads()\n",
|
||||
"model.dtype"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"id": "mcpm-16",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# LoRA 配置\n",
|
||||
"\n",
|
||||
"MiniCPM5-1B 是标准 Llama 架构，LoRA 目标模块与 Llama 一致。"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"id": "mcpm-17",
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"config = LoraConfig(\n",
|
||||
" task_type=TaskType.CAUSAL_LM,\n",
|
||||
" target_modules=[\n",
|
||||
" \"q_proj\", \"k_proj\", \"v_proj\", \"o_proj\",\n",
|
||||
" \"gate_proj\", \"up_proj\", \"down_proj\",\n",
|
||||
" ],\n",
|
||||
" inference_mode=False, # 训练模式\n",
|
||||
" r=8, # Lora 秩\n",
|
||||
" lora_alpha=32, # Lora alpha，缩放系数 = 32/8 = 4\n",
|
||||
" lora_dropout=0.1, # Dropout 比例\n",
|
||||
")\n",
|
||||
"config"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"id": "mcpm-18",
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"model = get_peft_model(model, config)\n",
|
||||
"model.print_trainable_parameters()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"id": "mcpm-19",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 配置训练参数"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"id": "mcpm-20",
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"args = TrainingArguments(\n",
|
||||
" output_dir=\"./output/MiniCPM5_1B_LoRA\",\n",
|
||||
" per_device_train_batch_size=4,\n",
|
||||
" gradient_accumulation_steps=4,\n",
|
||||
" logging_steps=10,\n",
|
||||
" num_train_epochs=3,\n",
|
||||
" save_steps=100,\n",
|
||||
" learning_rate=1e-4,\n",
|
||||
" save_on_each_node=True,\n",
|
||||
" gradient_checkpointing=True,\n",
|
||||
" report_to=\"none\",\n",
|
||||
")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"id": "mcpm-21",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# SwanLab 简介\n",
|
||||
"\n",
|
||||
"[SwanLab](https://github.com/swanhubx/swanlab) 是一个开源的模型训练记录工具，提供训练可视化、自动日志记录、超参数记录、实验对比、多人协同等功能。\n",
|
||||
"\n",
|
||||
"建议先在 [SwanLab 官网](https://swanlab.cn/) 注册账号，初始化时选择 `(2) Use an existing SwanLab account` 并使用 private API Key 登录。"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"id": "mcpm-22",
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"# 实例化 SwanLabCallback\n",
|
||||
"# 首次使用会提示登录，输入你在 SwanLab 官网获取的 API Key\n",
|
||||
"swanlab_callback = SwanLabCallback(\n",
|
||||
" project=\"MiniCPM5-Lora\",\n",
|
||||
" experiment_name=\"MiniCPM5-1B-LoRA\",\n",
|
||||
")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"id": "mcpm-23",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 训练"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"id": "mcpm-24",
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"trainer = Trainer(\n",
|
||||
" model=model,\n",
|
||||
" args=args,\n",
|
||||
" train_dataset=tokenized_id,\n",
|
||||
" data_collator=DataCollatorForSeq2Seq(tokenizer=tokenizer, padding=True),\n",
|
||||
" callbacks=[swanlab_callback],\n",
|
||||
")\n",
|
||||
"\n",
|
||||
"trainer.train()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"id": "mcpm-25",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 加载 LoRA 权重推理\n",
|
||||
"\n",
|
||||
"得到 checkpoint 后，加载基础模型并挂载 LoRA 权重进行推理。"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"id": "mcpm-26",
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"from peft import PeftModel\n",
|
||||
"\n",
|
||||
"# 基础模型路径 & 训练得到的 LoRA 权重路径（按实际 checkpoint 编号修改）\n",
|
||||
"lora_path = \"./output/MiniCPM5_1B_LoRA/checkpoint-702\"\n",
|
||||
"\n",
|
||||
"# 加载基础模型\n",
|
||||
"base_model = AutoModelForCausalLM.from_pretrained(\n",
|
||||
" model_id, dtype=torch.bfloat16, device_map=\"auto\"\n",
|
||||
")\n",
|
||||
"# 挂载 LoRA 权重\n",
|
||||
"model = PeftModel.from_pretrained(base_model, model_id=lora_path)\n",
|
||||
"model.eval()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"id": "mcpm-27",
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"prompt = \"你是谁？\"\n",
|
||||
"messages = [\n",
|
||||
" {\"role\": \"system\", \"content\": \"现在你要扮演皇帝身边的女人--甄嬛\"},\n",
|
||||
" {\"role\": \"user\", \"content\": prompt},\n",
|
||||
"]\n",
|
||||
"inputs = tokenizer.apply_chat_template(\n",
|
||||
" messages,\n",
|
||||
" add_generation_prompt=True,\n",
|
||||
" enable_thinking=False, # 关闭思考模式，直接输出角色回答\n",
|
||||
" tokenize=True,\n",
|
||||
" return_dict=True,\n",
|
||||
" return_tensors=\"pt\",\n",
|
||||
").to(model.device)\n",
|
||||
"\n",
|
||||
"gen_kwargs = {\"max_new_tokens\": 128, \"do_sample\": True, \"top_p\": 0.95, \"temperature\": 0.7}\n",
|
||||
"with torch.no_grad():\n",
|
||||
" outputs = model.generate(**inputs, **gen_kwargs)\n",
|
||||
"outputs = outputs[:, inputs[\"input_ids\"].shape[1]:]\n",
|
||||
"print(tokenizer.decode(outputs[0], skip_special_tokens=True))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"id": "2e2cc414-c5fb-43ff-a4c1-9ab8c27bd03b",
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": []
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3 (ipykernel)",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.12.3"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 5
|
||||
}
|
||||
@@ -0,0 +1,317 @@
|
||||
# MiniCPM5-1B-LoRA 及 SwanLab 可视化记录
|
||||
|
||||
> 本教程配套 notebook：[03-MiniCPM5-1B-LoRA.ipynb](./03-MiniCPM5-1B-LoRA.ipynb)
|
||||
|
||||
## MiniCPM5-1B 简介
|
||||
|
||||
`MiniCPM5-1B` 是面壁智能（ModelBest）/ OpenBMB 发布的 1B 稠密 Transformer，采用**标准 `LlamaForCausalLM` 架构**（24 层，GQA，128K 上下文）。它内置 `<think>` chat template，支持「思考 / 非思考」双模式（`enable_thinking` 切换），并原生支持工具调用。1B 的体量非常适合在单卡上做 LoRA 微调实验。
|
||||
|
||||
本教程使用官方推荐的纯 `transformers + peft` 方案完成 LoRA 微调，并使用 **SwanLab** 记录训练过程。
|
||||
|
||||
## 环境配置
|
||||
|
||||
```bash
|
||||
# 换清华镜像源
|
||||
pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
|
||||
|
||||
# 核心依赖（MiniCPM5 需要 transformers>=5.6）
|
||||
pip install "transformers>=5.6"
|
||||
pip install accelerate datasets peft swanlab modelscope
|
||||
```
|
||||
|
||||
> 考虑到部分同学配置环境可能会遇到一些问题，我们在 AutoDL 平台准备了环境镜像，点击下方链接并直接创建 Autodl 示例即可。
|
||||
> ***https://www.codewithgpu.com/i/datawhalechina/self-llm/MiniCPM5***
|
||||
|
||||
## 模型下载
|
||||
|
||||
```python
|
||||
# model_download.py
|
||||
from modelscope import snapshot_download
|
||||
|
||||
model_dir = snapshot_download('OpenBMB/MiniCPM5-1B', cache_dir='/root/autodl-tmp')
|
||||
print(f"模型下载完成，保存路径为：{model_dir}")
|
||||
```
|
||||
|
||||
然后在终端中输入 `python model_download.py` 执行下载。
|
||||
|
||||
> 注意：记得修改 `cache_dir` 为你的模型下载路径哦~
|
||||
|
||||
## 数据集构建
|
||||
|
||||
对大语言模型进行 `supervised-finetuning`（`sft`，有监督微调）的数据格式如下：
|
||||
|
||||
```json
|
||||
{
|
||||
"instruction": "回答以下用户问题，仅输出答案。",
|
||||
"input": "1+1等于几?",
|
||||
"output": "2"
|
||||
}
|
||||
```
|
||||
|
||||
其中，`instruction` 是用户指令；`input` 是用户输入；`output` 是模型应该给出的输出。
|
||||
|
||||
我们的目标是通过大量人物对话数据微调得到一个能够 role-play 甄嬛对话风格的模型，数据示例如下：
|
||||
|
||||
```json
|
||||
{
|
||||
"instruction": "你父亲是谁？",
|
||||
"input": "",
|
||||
"output": "家父是大理寺少卿甄远道。"
|
||||
}
|
||||
```
|
||||
|
||||
本教程使用的甄嬛对话示例微调数据集位于 [/dataset/huanhuan.json](../../dataset/huanhuan.json)（共 3729 条），数据格式为 `instruction / input / output` 的 Alpaca 格式。
|
||||
|
||||
## 数据准备
|
||||
|
||||
LoRA 训练的数据需要经过格式化、编码之后再输入给模型。这里我们直接使用 tokenizer 自带的 `apply_chat_template` 构造对话模板。
|
||||
|
||||
### 认识 MiniCPM5 的 Chat Template
|
||||
|
||||
`MiniCPM5-1B` 采用 `<|im_start|>role\n...<|im_end|>\n` 格式，并支持 `enable_thinking` 参数控制思考模式。对于「角色扮演」任务，我们关闭思考模式（`enable_thinking=False`）：
|
||||
|
||||
```python
|
||||
from transformers import AutoTokenizer
|
||||
|
||||
model_id = '/root/autodl-tmp/OpenBMB/MiniCPM5-1B'
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||||
|
||||
messages = [
|
||||
{"role": "system", "content": "现在你要扮演皇帝身边的女人--甄嬛"},
|
||||
{"role": "user", "content": "你父亲是谁？"},
|
||||
{"role": "assistant", "content": "家父是大理寺少卿甄远道。"},
|
||||
]
|
||||
|
||||
text = tokenizer.apply_chat_template(
|
||||
messages, tokenize=False, add_generation_prompt=False, enable_thinking=False
|
||||
)
|
||||
print(text)
|
||||
```
|
||||
|
||||
输出如下：
|
||||
|
||||
```
|
||||
<s><|im_start|>system
|
||||
现在你要扮演皇帝身边的女人--甄嬛<|im_end|>
|
||||
<|im_start|>user
|
||||
你父亲是谁？<|im_end|>
|
||||
<|im_start|>assistant
|
||||
家父是大理寺少卿甄远道。<|im_end|>
|
||||
```
|
||||
|
||||
### 构造处理函数
|
||||
|
||||
> **注意一个细节**：MiniCPM5 的模板在「带 generation prompt」时会追加 `<think>\n\n</think>\n\n`（非思考模式占位），但「完整对话渲染」时助手回合并**不**包含这个 think 块。因此这里**不能用 token 级切片**（`full[len(prompt):]`），而要分别对「前缀」和「回答」单独 tokenize 再拼接。
|
||||
|
||||
```python
|
||||
def process_func(example):
|
||||
MAX_LENGTH = 1024
|
||||
SYS = "现在你要扮演皇帝身边的女人--甄嬛"
|
||||
|
||||
messages = [{"role": "system", "content": SYS},
|
||||
{"role": "user", "content": example["instruction"] + example["input"]}]
|
||||
# 前缀（system + user，带 generation prompt，含非思考 think 占位），不计算 loss
|
||||
prompt_ids = tokenizer.apply_chat_template(
|
||||
messages, tokenize=True, add_generation_prompt=True,
|
||||
enable_thinking=False, return_dict=False,
|
||||
)
|
||||
# 回答部分：output + 结束符 <|im_end|>
|
||||
response_ids = tokenizer(example["output"], add_special_tokens=False).input_ids \
|
||||
+ [tokenizer.convert_tokens_to_ids("<|im_end|>")]
|
||||
|
||||
input_ids = prompt_ids + response_ids
|
||||
labels = [-100] * len(prompt_ids) + response_ids
|
||||
attention_mask = [1] * len(input_ids)
|
||||
|
||||
if len(input_ids) > MAX_LENGTH: # 超长截断
|
||||
input_ids = input_ids[:MAX_LENGTH]
|
||||
attention_mask = attention_mask[:MAX_LENGTH]
|
||||
labels = labels[:MAX_LENGTH]
|
||||
|
||||
return {"input_ids": input_ids, "attention_mask": attention_mask, "labels": labels}
|
||||
```
|
||||
|
||||
> 说明：MiniCPM5 的 `<|im_end|>` token id 为 `130073`，`</s>`（eos）为 `1`。这里用 `<|im_end|>` 作为回答的结束符，与模板一致。
|
||||
|
||||
读入数据集并应用处理函数：
|
||||
|
||||
```python
|
||||
import json
|
||||
from datasets import Dataset
|
||||
|
||||
with open("/root/autodl-tmp/huanhuan.json", "r", encoding="utf-8") as f:
|
||||
data = json.load(f)
|
||||
|
||||
ds = Dataset.from_list(data)
|
||||
tokenized_id = ds.map(process_func, remove_columns=ds.column_names)
|
||||
tokenized_id
|
||||
```
|
||||
|
||||
可以解码查看处理后的样本：
|
||||
|
||||
```python
|
||||
print(tokenizer.decode(tokenized_id[0]["input_ids"]))
|
||||
print(tokenizer.decode(list(filter(lambda x: x != -100, tokenized_id[0]["labels"]))))
|
||||
```
|
||||
|
||||
```
|
||||
<s><|im_start|>system
|
||||
现在你要扮演皇帝身边的女人--甄嬛<|im_end|>
|
||||
<|im_start|>user
|
||||
小姐，别的秀女都在求中选，唯有咱们小姐想被撂牌子，菩萨一定记得真真儿的——<|im_end|>
|
||||
<|im_start|>assistant
|
||||
<think>
|
||||
|
||||
</think>
|
||||
|
||||
嘘——都说许愿说破是不灵的。<|im_end|>
|
||||
```
|
||||
|
||||
`labels`（过滤掉 -100）：
|
||||
|
||||
```
|
||||
嘘——都说许愿说破是不灵的。<|im_end|>
|
||||
```
|
||||
|
||||
## 加载模型和 tokenizer
|
||||
|
||||
`MiniCPM5-1B` 是标准 `LlamaForCausalLM`，直接用 `AutoModelForCausalLM` 加载：
|
||||
|
||||
```python
|
||||
import torch
|
||||
from transformers import AutoModelForCausalLM
|
||||
|
||||
model = AutoModelForCausalLM.from_pretrained(
|
||||
'/root/autodl-tmp/OpenBMB/MiniCPM5-1B',
|
||||
dtype=torch.bfloat16,
|
||||
device_map="auto",
|
||||
)
|
||||
model.enable_input_require_grads() # 开启梯度检查点时需要
|
||||
model.dtype # torch.bfloat16
|
||||
```
|
||||
|
||||
## LoRA Config
|
||||
|
||||
`MiniCPM5-1B` 是标准 Llama 架构，LoRA 目标模块与 Llama 一致：`q_proj / k_proj / v_proj / o_proj / gate_proj / up_proj / down_proj`。
|
||||
|
||||
```python
|
||||
from peft import LoraConfig, TaskType, get_peft_model
|
||||
|
||||
config = LoraConfig(
|
||||
task_type=TaskType.CAUSAL_LM,
|
||||
target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
|
||||
inference_mode=False, # 训练模式
|
||||
r=8, # Lora 秩
|
||||
lora_alpha=32, # Lora alpha，缩放系数 = 32/8 = 4
|
||||
lora_dropout=0.1, # Dropout 比例
|
||||
)
|
||||
|
||||
model = get_peft_model(model, config)
|
||||
model.print_trainable_parameters()
|
||||
```
|
||||
|
||||
输出（仅训练约 0.5% 的参数）：
|
||||
|
||||
```
|
||||
trainable params: 5,603,328 || all params: 1,086,236,160 || trainable%: 0.5158
|
||||
```
|
||||
|
||||
## Training Arguments
|
||||
|
||||
```python
|
||||
from transformers import TrainingArguments, Trainer, DataCollatorForSeq2Seq
|
||||
|
||||
args = TrainingArguments(
|
||||
output_dir="./output/MiniCPM5_1B_LoRA",
|
||||
per_device_train_batch_size=4,
|
||||
gradient_accumulation_steps=4,
|
||||
logging_steps=10,
|
||||
num_train_epochs=3,
|
||||
save_steps=100,
|
||||
learning_rate=1e-4,
|
||||
save_on_each_node=True,
|
||||
gradient_checkpointing=True,
|
||||
report_to="none",
|
||||
)
|
||||
```
|
||||
|
||||
## SwanLab 简介
|
||||
|
||||
|
||||
|
||||
[SwanLab](https://github.com/swanhubx/swanlab) 是一个开源的模型训练记录工具，提供训练可视化、自动日志记录、超参数记录、实验对比、多人协同等功能。
|
||||
|
||||
**为什么要记录训练**：模型训练更像一门实验科学，一个优秀模型背后往往是成千上万次实验。高效记录与对比对研究效率至关重要。
|
||||
|
||||
## 实例化 SwanLabCallback
|
||||
|
||||
建议先在 [SwanLab 官网](https://swanlab.cn/) 注册账号，初始化时选择 `(2) Use an existing SwanLab account` 并使用 private API Key 登录。
|
||||
|
||||
```python
|
||||
import swanlab
|
||||
from swanlab.integration.transformers import SwanLabCallback
|
||||
|
||||
swanlab_callback = SwanLabCallback(
|
||||
project="MiniCPM5-Lora",
|
||||
experiment_name="MiniCPM5-1B-LoRA",
|
||||
)
|
||||
```
|
||||
|
||||
## 使用 Trainer 训练
|
||||
|
||||
```python
|
||||
trainer = Trainer(
|
||||
model=model,
|
||||
args=args,
|
||||
train_dataset=tokenized_id,
|
||||
data_collator=DataCollatorForSeq2Seq(tokenizer=tokenizer, padding=True),
|
||||
callbacks=[swanlab_callback],
|
||||
)
|
||||
|
||||
trainer.train()
|
||||
```
|
||||
|
||||
训练完成后，打开 SwanLab 即可查看训练过程中记录的参数与 loss 曲线。
|
||||
|
||||
## 加载 LoRA 权重推理
|
||||
|
||||
```python
|
||||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||||
import torch
|
||||
from peft import PeftModel
|
||||
|
||||
model_id = '/root/autodl-tmp/OpenBMB/MiniCPM5-1B'
|
||||
lora_path = './output/MiniCPM5_1B_LoRA/checkpoint-XXX' # 按实际 checkpoint 填写
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||||
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto")
|
||||
model = PeftModel.from_pretrained(model, model_id=lora_path)
|
||||
model.eval()
|
||||
|
||||
messages = [
|
||||
{"role": "system", "content": "现在你要扮演皇帝身边的女人--甄嬛"},
|
||||
{"role": "user", "content": "你是谁？"},
|
||||
]
|
||||
inputs = tokenizer.apply_chat_template(
|
||||
messages,
|
||||
add_generation_prompt=True,
|
||||
enable_thinking=False,
|
||||
tokenize=True,
|
||||
return_dict=True,
|
||||
return_tensors="pt",
|
||||
).to(model.device)
|
||||
|
||||
gen_kwargs = {"max_new_tokens": 128, "do_sample": True, "top_p": 0.95, "temperature": 0.7}
|
||||
with torch.no_grad():
|
||||
outputs = model.generate(**inputs, **gen_kwargs)
|
||||
outputs = outputs[:, inputs["input_ids"].shape[1]:]
|
||||
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
||||
```
|
||||
|
||||
输出示例：
|
||||
|
||||
```
|
||||
我是甄嬛，家父是大理寺少卿甄远道。
|
||||
```
|
||||
|
||||
可以看到，经过 LoRA 微调后，模型已经学会了甄嬛的说话风格与人物设定。
|
||||
Binary file not shown.
|
After Width: | Height: | Size: 36 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 63 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 28 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 34 KiB |
@@ -4,6 +4,7 @@
|
||||
|
||||
## 目录
|
||||
|
||||
- [MiniCPM5](#minicpm5)
|
||||
- [Kimi-K2.5](#kimi-k25)
|
||||
- [Step-3.5-Flash](#step-35-flash)
|
||||
- [GLM-4.7-Flash](#glm-47-flash)
|
||||
@@ -62,6 +63,14 @@
|
||||
|
||||
## 已支持模型列表
|
||||
|
||||
### MiniCPM5
|
||||
|
||||
[MiniCPM5-1B](https://www.modelscope.cn/models/OpenBMB/MiniCPM5-1B)
|
||||
- [x] [MiniCPM5-1B vLLM 部署调用](./models/minicpm5/01-MiniCPM5-1B-vLLM%20部署调用.md)
|
||||
- [x] [MiniCPM5-1B SGLang 部署调用](./models/minicpm5/02-MiniCPM5-1B-SGLang%20部署调用.md)
|
||||
- [x] [MiniCPM5-1B LoRA 及 SwanLab 可视化记录](./models/minicpm5/03-MiniCPM5-1B-LoRA及SwanLab可视化记录.md)
|
||||
- [x] [MiniCPM5-1B LoRA Docker 镜像](https://www.codewithgpu.com/i/datawhalechina/self-llm/MiniCPM5)
|
||||
|
||||
### Kimi-K2.5
|
||||
|
||||
[Kimi-K2.5](https://huggingface.co/moonshotai/Kimi-K2.5)
|
||||
|
||||
Reference in New Issue
Block a user