Yuan Lab袁静课题组

The first step of deep-learning development is a clean, reproducible environment. This guide covers the standard approach: create an isolated environment with conda, install a PyTorch build matching your GPU, and verify that CUDA works.

Why an environment manager

Different projects need different PyTorch / CUDA versions; installing into the system Python eventually conflicts. conda (or venv) gives each project its own environment that can be removed cleanly. Miniconda is a lightweight starting point.

Create an environment

conda create -n dl python=3.10 -y
conda activate dl

Follow the project README for the Python version; 3.10 is a safe default today.

Install PyTorch

The key is matching the PyTorch CUDA build to your driver. Check the driver ceiling first:

nvidia-smi   # the CUDA Version at top-right is the max the driver supports

Then pick the matching command from the PyTorch website, e.g.:

# CUDA 12.1 example (always use the command from the official site)
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

# CPU-only machines
pip install torch torchvision

Verify the install

python -c "import torch; print(torch.__version__, torch.cuda.is_available())"

A PyTorch version plus cuda.is_available() returning True means success.

FAQ

• cuda.is_available() is False: the build and driver mismatch (old driver, or a CPU build). Recheck nvidia-smi and reinstall the matching build.
• Slow downloads: configure a local pip mirror or use conda channels.
• Dependency conflicts across projects: one environment per project; never share.

For lab-server quotas, queues, and mirror configuration, see the lab’s internal notes or ask the administrator.