Install#

Standard tutorial path#

For the primary documentation and notebook workflow, install scDLKit with the tutorial extras:

python -m pip install "scdlkit[tutorials]"

On Windows, prefer a short virtual-environment path such as C:\venvs\scdlkit if you install the tutorials extra. The bundled Jupyter dependencies can exceed Windows path-length limits when the environment path is deeply nested.

That installs:

  • scanpy

  • jupyter

  • the core scDLKit package

The public tutorials default to a quickstart profile. Each notebook also includes a full profile for longer, more convincing qualitative runs without changing the overall workflow.

CPU and GPU#

scDLKit uses the same tutorial code on CPU and GPU. The notebooks and scripts should default to device="auto", so the package will use CUDA when available and fall back to CPU otherwise.

CPU or default install#

python -m pip install "scdlkit[tutorials]"

GPU install#

Install the matching PyTorch build for your platform first, then install the tutorial extras:

python -m pip install "scdlkit[tutorials]"

Use the official PyTorch install selector for the correct CUDA or accelerator-specific command:

Minimal package install#

If you only want the core library without Scanpy or notebooks:

python -m pip install scdlkit

Available extras:

  • scdlkit[notebook]

  • scdlkit[scanpy]

  • scdlkit[tutorials]

  • scdlkit[dev]

  • scdlkit[docs]

First command to run#

Most users should open the Scanpy-first quickstart notebook:

jupyter notebook examples/train_vae_pbmc.ipynb

Then keep the same notebook and the same device="auto" setting for CPU or GPU. If you want a longer run, change the first config cell from TUTORIAL_PROFILE = "quickstart" to TUTORIAL_PROFILE = "full".