step: install_torch
description: "安装 PyTorch(含 CUDA 支持验证)"
validation:
script: |
python3 -c "import torch; assert torch.cuda.is_available(); print(torch.__version__, torch.cuda.get_device_capability(0))"
expected_pattern: "({gpu_cc})"
branches:
- id: pip_nvidia_index
description: "pip install torch(NVIDIA CUDA index,自动匹配系统 CUDA 版本)"
command_template: "pip install torch --index-url https://download.pytorch.org/whl/cu{cuda_ver_short}"
fallback_on: ["driver too old", "CUDA driver version is insufficient"]
note: |
- 从 nvidia-smi 自动检测 CUDA 版本(如 12.8 → cu128, 12.4 → cu124)
- 国内需配合阿里云镜像加速
- id: pip_nvidia_index_mirror
description: "pip install torch(NVIDIA CUDA index + 阿里云镜像加速)"
command_template: "pip install torch --index-url https://download.pytorch.org/whl/cu{cuda_ver_short} -i https://mirrors.aliyun.com/pypi/simple/"
note: "阿里云镜像 aarch64 wheel 覆盖可能不全"
- id: pip_default
description: "pip install torch(默认 PyPI,aarch64 上通常给 CPU-only)"
command: "pip install torch"
fallback_on: ["No matching distribution", "CPU-only", "ModuleNotFoundError: No module named 'torch'"]
- id: conda
description: "conda install pytorch(自动匹配 CUDA 版本)"
command_template: "conda install pytorch torchvision torchaudio pytorch-cuda={cuda_ver} -c pytorch -c nvidia"
fallback_on: ["conda: command not found", "channel not found"]
note: "aarch64 conda channel 支持有限"