README.md (2053B)
1 # Prepare Python 2 ## create venv 3 ### install uv 4 `pip install uv` 5 ### create .venv 6 `uv venv .venv` 7 ## activate venv 8 `source .venv/bin/activate` 9 if you use windows powershell, use `.\.venv\Scripts\activate.ps1` 10 11 ## install requirements 12 `uv pip install -r requirements.txt` 13 14 ## setting for cuda 15 please make sure you have cuda installed and set the environment variable CUDA_HOME to the path of your cuda installation. 16 check for it can do with `nvcc --version` 17 and install pytorch for your cuda version. 18 for example nvcc version is 13.0 19 `uv pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130 --reinstall` 20 And then, check this python env can use cuda with `python gpu_check.py` 21 22 ## set API key for gemini(only need to use structuring-training-data) 23 `export GEMINI_API_KEY=your_api_key` 24 if you use windows powershell, use `$env:GEMINI_API_KEY="あなたのAPIキー"` 25 26 # Prepare Feedback 27 you only need to use `python structuring-training-data.py` 28 29 # integrate data 30 integrate feedback and report's typst 31 use `integrate_data.py` 32 33 # training model 34 use `train.py` 35 it may cost many time. 36 My RTX 3060 spend 10min. 37 38 # convert model to gguf 39 ``` 40 python vendor/llama.cpp/convert_hf_to_gguf.py mistral_typst_merged --outfile mistral_typst_merged.f16.gguf --outtype f16 41 ``` 42 # quantize model 43 ``` 44 python quantize_gguf.py mistral_typst_merged.f16.gguf mistral_typst_merged.q4_0.gguf --quantize-fn q4_0 45 ``` 46 なおこのプログラムにはllama.cppのビルド済みバイナリが配置されているので動作している． 47 必要に応じて最新のビルド済みバイナリに更新したり，自分でビルドしてください． 48 49 # サーバーの起動 50 ``` 51 uvicorn api_server:app --host 0.0.0.0 --port 8080 52 ``` 53 サーバーがcudaを使わないときは，これを参考にしましたが，治りません．dll地獄です． 54 https://zenn.dev/hellohazime/articles/ccd01c2df0b5c3 55 56 # レポートのレビュー 57 use `review_typst.py` 58 it may cost many time.