PyTorch Poetry Boilerplate (CUDA + CPU)

Language: English | Русский

Ready-to-use pyproject.toml templates for PyTorch-based neural network projects, optimized for both local development and future PyPI distribution.

---

Files

• pyproject-cuda128.toml — installs PyTorch with CUDA 12.8 support from the official PyTorch wheel index.
• pyproject-cpu.toml — pure CPU-only version using standard PyPI wheels.

Both files are pre-configured with package-mode = false for local experimentation (no need to structure your code as a package).

---

Need to change CUDA version?

• Define which version of CUDA your GPU supports. You can do it here: CUDA Compability Table
• Make sure that CUDA installed on your PC. Guide: CUDA Quick Start | Windows | Linux
• In pyproject-cuda128.toml change source url = "https://download.pytorch.org/whl/cu128" to your version. For example, for CUDA v. 12.6 the end of url will be cu126

Quick Start

1. Choose your version:

   # For NVIDIA GPU (CUDA 12.8)
   cp pyproject-cuda128.toml pyproject.toml
   
   # OR for CPU-only
   cp pyproject-cpu.toml pyproject.toml
   

2. Initialize Poetry environment:

poetry install

3. Run poetry env activate

4. Start coding! No src/ layout or package structure required

---

Want publish to PyPi Later?

1. Set package-mode = true
2. Uncomment this line:

# packages = [{ include = "your_package_name", from = "src" }]

3. Move your code into src/your_package_name/
4. Ensure your project has a valid __init__.py
5. Run poetry build and poetry publish

Keep package-mode = false during research/prototyping. Switch to true only when preparing a distributable library

---

Included Dependencies

• torch, torchaudio, torchvision (CUDA or CPU)
• tensorboard
• numpy, scikit-learn
• matplotlib, seaborn, pandas
• tqdm, pyyaml

---

Don't need some of included dependencies?

Just remove it:

poetry remove unnecessary-dependency-name

License

MIT — feel free to fork, modify, and use in your projects.