🧠 两小时从 0 训出你的 64M 大模型。Train your own 64M LLM from scratch in 2 hours.
[feat] update readme
jingyaogongeb96113cd4ac
diff --git a/README.md b/README.mdindex bfa0614..3fcf61f 100644--- a/README.md+++ b/README.md@@ -1659,9 +1659,9 @@ MiniMind模型本身预训练数据集小的可怜,也没有针对性的对测## 模型转换-* [./scripts/convert_model.py](./scripts/convert_model.py)可以实现`torch模型/transformers`模型之间的转换+* [./scripts/convert_model.py](./scripts/convert_model.py)可以实现`torch / transformers`模型的互相转换+* 如无特别说明,`MiniMind2`模型均默认为`Transformers`格式的模型,需提前`t2t`转换!----## 基于MiniMind-API服务接口@@ -1670,13 +1670,12 @@ MiniMind模型本身预训练数据集小的可怜,也没有针对性的对测* 从[Huggingface](https://huggingface.co/collections/jingyaogong/minimind-66caf8d999f5c7fa64f399e5)下载模型权重文件,文件树:```- <MiniMind-Model-Name> (root dir)- ├─<MiniMind-Model-Name>+ minimind (root dir)+ ├─<MiniMind-Model-Name>(例如MiniMind2)| ├── config.json| ├── generation_config.json- | ├── LMConfig.py- | ├── model.py- | ├── pytorch_model.bin+ | ├── model_minimind.py or w/o+ | ├── pytorch_model.bin or model.safetensors| ├── special_tokens_map.json| ├── tokenizer_config.json| ├── tokenizer.json@@ -1709,6 +1708,8 @@ MiniMind模型本身预训练数据集小的可怜,也没有针对性的对测vLLM是极其流行的高效推理框架,支持大模型快速部署,优化显存利用与吞吐量。+以openai-serve形式启动 minimind2:+```bashvllm serve ./MiniMind2 --model-impl transformers --served-model-name "minimind" --port 8998```@@ -1718,59 +1719,106 @@ vllm serve ./MiniMind2 --model-impl transformers --served-model-name "minimind"llama.cpp是一个C++库,可以在命令行下直接使用,支持多线程推理,支持GPU加速。-参考官方仓库安装后,在`convert_hf_to_gguf.py` ~760行插入+**目录结构**:建议将llama.cpp与minimind放在同级目录下-```text-# 添加MiniMind2 tokenizer支持+```+parent/+├── minimind/ # MiniMind项目目录+│ ├── MiniMind2/ # HuggingFace格式MiniMind2模型 (先convert_model.py生成)+│ │ ├── config.json+│ │ ├── model.safetensors+│ │ └── ...+│ ├── model/+│ ├── trainer/+│ └── ...+└── llama.cpp/ # llama.cpp项目目录+ ├── build/+ ├── convert_hf_to_gguf.py+ └── ...+```++0、参考`llama.cpp`官方步骤进行install++1、在`convert_hf_to_gguf.py`的`get_vocab_base_pre`函数最后插入:++```python+# 添加MiniMind tokenizer支持(这里随便写一个例如qwen2即可)if res is None:- res = "smollm"+ res = "qwen2"```-转换自定义训练的minimind模型 -> gguf+2、转换自训练的minimind模型:huggingface -> gguf```bash-python convert_hf_to_gguf.py ../minimind/MiniMind2/+# 在llama.cpp下执行,将生成../minimind/MiniMind2/MiniMind2-xxx.gguf+python convert_hf_to_gguf.py ../minimind/MiniMind2```-量化模型+3、量化此模型 (可选)```bash-./build/bin/llama-quantize ../minimind/MiniMind2/MiniMind2-109M-F16.gguf ../minimind/MiniMind2/Q4-MiniMind2.gguf Q4_K_M+./build/bin/llama-quantize ../minimind/MiniMind2/MiniMind2.gguf ../minimind/MiniMind2/Q4-MiniMind2.gguf Q4_K_M```-命令行推理+4、命令行推理测试```bash-./build/bin/llama-cli -m ../minimind/MiniMind2/MiniMind2-109M-F16.gguf --chat-template chatml+./build/bin/llama-cli -m ../minimind/MiniMind2/MiniMind2.gguf -sys "You are a helpful assistant" # system prompt必须固定```## <img src="https://ollama.com/public/cloud.png" height="28" style="vertical-align: middle;"/> [ollama](https://ollama.ai)ollama是本地运行大模型的工具,支持多种开源LLM,简单易用。-通过ollama加载自定义的gguf模型,新建minimind.modelfile:+1、通过ollama加载自定义的gguf模型++在`MiniMind2`下新建`minimind.modelfile`,写入:```text-FROM ./MiniMind2-109M-F16.gguf-TEMPLATE """{{ if .System }}<|im_start|>system+FROM ./Q4-MiniMind2.gguf++SYSTEM """You are a helpful assistant"""++TEMPLATE """<|im_start|>system{{ .System }}<|im_end|>-{{ end }}{{ if .Prompt }}<|im_start|>user+<|im_start|>user{{ .Prompt }}<|im_end|>-{{ end }}<|im_start|>assistant+<|im_start|>assistant+{{ .Response }}<|im_end|>"""```-加载模型并命名为`minimind2`+2、加载并命名此模型为`minimind-local````bash-ollama create -f minimind.modelfile minimind2+ollama create -f minimind.modelfile minimind-local```-启动推理+3、启动推理-```text-ollama run minimind2-> 你好,我是MiniMind2,一个基于xxxxxxxx+```bash+ollama run minimind-local+```++<details>+<summary>📤 推送你的模型到 Ollama Hub</summary>++```bash+# 1. 为本地模型重命名为你的ollama-account/minimind的tag+ollama cp minimind-local:latest your_username/minimind:latest++# 2. 推送模型+ollama push your_username/minimind:latest+```+</details>+<br/>++⭐️ 也可直接使用我提供的ollama模型一键启动:++```bash+ollama run jingyaogong/minimind2 # 其他可选 minimind2-r1 / minimind2-small / minimind2-small-r1+>>> 你叫什么名字+我是一个语言模型...```> 以上三方框架的更多用法请参考对应官方文档😊diff --git a/README_en.md b/README_en.mdindex 012ab60..67081b5 100644--- a/README_en.md+++ b/README_en.md@@ -1582,9 +1582,10 @@ Models generally achieve baseline performance due to small parameter scales and## Model Conversion-* [./scripts/convert_model.py](./scripts/convert_model.py) enables conversion between `torch models/transformers` models+* [./scripts/convert_model.py](./scripts/convert_model.py) enables mutual conversion of `torch / transformers` models+* Unless otherwise specified, `MiniMind2` models are by default in `Transformers` format and require `t2t` conversion beforehand!+----## OpenAI-API Based MiniMind Service Interface@@ -1592,13 +1593,12 @@ Models generally achieve baseline performance due to small parameter scales and* Download model weights from [Huggingface](https://huggingface.co/collections/jingyaogong/minimind-66caf8d999f5c7fa64f399e5), file structure:```- <MiniMind-Model-Name> (root dir)- ├─<MiniMind-Model-Name>+ minimind (root dir)+ ├─<MiniMind-Model-Name> (e.g. MiniMind2)| ├── config.json| ├── generation_config.json- | ├── LMConfig.py- | ├── model.py- | ├── pytorch_model.bin+ | ├── model_minimind.py or w/o+ | ├── pytorch_model.bin or model.safetensors| ├── special_tokens_map.json| ├── tokenizer_config.json| ├── tokenizer.json@@ -1631,6 +1631,8 @@ Models generally achieve baseline performance due to small parameter scales andvLLM is an extremely popular efficient inference framework supporting fast deployment of large models, optimizing GPU memory usage and throughput.+Start minimind2 in openai-serve format:+```bashvllm serve ./MiniMind2 --model-impl transformers --served-model-name "minimind" --port 8998```@@ -1639,62 +1641,109 @@ vllm serve ./MiniMind2 --model-impl transformers --served-model-name "minimind"llama.cpp is a C++ library that can be used directly from command line, supports multi-threaded inference, and supports GPU acceleration.-After referencing official repository installation, insert at ~760 line in `convert_hf_to_gguf.py`+**Directory Structure**: It is recommended to place llama.cpp and minimind in the same parent directory-```text-# Add MiniMind2 tokenizer support+```+parent/+├── minimind/ # MiniMind project directory+│ ├── MiniMind2/ # HuggingFace format MiniMind2 model (generated by convert_model.py first)+│ │ ├── config.json+│ │ ├── model.safetensors+│ │ └── ...+│ ├── model/+│ ├── trainer/+│ └── ...+└── llama.cpp/ # llama.cpp project directory+ ├── build/+ ├── convert_hf_to_gguf.py+ └── ...+```++0. Follow the official `llama.cpp` installation steps++1. Insert at the end of the `get_vocab_base_pre` function in `convert_hf_to_gguf.py`:++```python+# Add MiniMind tokenizer support (you can use any existing one like qwen2)if res is None:- res = "smollm"+ res = "qwen2"```-Convert custom-trained minimind model -> gguf+2. Convert your custom-trained minimind model: huggingface -> gguf```bash-python convert_hf_to_gguf.py ../minimind/MiniMind2/+# Execute under llama.cpp, will generate ../minimind/MiniMind2/MiniMind2-xxx.gguf+python convert_hf_to_gguf.py ../minimind/MiniMind2```-Quantize model+3. Quantize the model (optional)```bash-./build/bin/llama-quantize ../minimind/MiniMind2/MiniMind2-109M-F16.gguf ../minimind/MiniMind2/Q4-MiniMind2.gguf Q4_K_M+./build/bin/llama-quantize ../minimind/MiniMind2/MiniMind2.gguf ../minimind/MiniMind2/Q4-MiniMind2.gguf Q4_K_M```-Command line inference+4. Command line inference test```bash-./build/bin/llama-cli -m ../minimind/MiniMind2/MiniMind2-109M-F16.gguf --chat-template chatml+./build/bin/llama-cli -m ../minimind/MiniMind2/MiniMind2.gguf -sys "You are a helpful assistant" # system prompt must be fixed```## <img src="https://ollama.com/public/cloud.png" height="28" style="vertical-align: middle;"/> [ollama](https://ollama.ai)ollama is a tool for running large models locally, supports multiple open-source LLMs, simple and easy to use.-Load custom gguf model through ollama, create new minimind.modelfile:+1. Load custom gguf model through ollama++Create `minimind.modelfile` under `MiniMind2`:```text-FROM ./MiniMind2-109M-F16.gguf-TEMPLATE """{{ if .System }}<|im_start|>system+FROM ./Q4-MiniMind2.gguf++SYSTEM """You are a helpful assistant"""++TEMPLATE """<|im_start|>system{{ .System }}<|im_end|>-{{ end }}{{ if .Prompt }}<|im_start|>user+<|im_start|>user{{ .Prompt }}<|im_end|>-{{ end }}<|im_start|>assistant+<|im_start|>assistant+{{ .Response }}<|im_end|>"""```-Load model and name as `minimind2`+2. Load and name this model as `minimind-local````bash-ollama create -f minimind.modelfile minimind2+ollama create -f minimind.modelfile minimind-local```-Start inference+3. Start inference-```text-ollama run minimind2-> Hello, I am MiniMind2, a model based on xxxxxxxx+```bash+ollama run minimind-local+```++<details>+<summary>📤 Push your model to Ollama Hub</summary>++```bash+# 1. Rename your local model to your_username/minimind tag+ollama cp minimind-local:latest your_username/minimind:latest++# 2. Push the model+ollama push your_username/minimind:latest+```+</details>+<br/>++⭐️ You can also directly use the ollama model I provided with one command:++```bash+ollama run jingyaogong/minimind2 # Other options: minimind2-r1 / minimind2-small / minimind2-small-r1+>>> What's your name?+I am a language model...```-> For more usage of the above three-party frameworks, please refer to their official documentation 😊+> For more usage of the above third-party frameworks, please refer to their official documentation 😊# 📌 Acknowledge