pytorch-wheels
January 1, 0001
skyzh
mirror-help
pytorch-wheels 是 PyTorch 官方预构建轮子（wheels）软件源的镜像。
PyTorch 根据不同的计算平台（CPU / CUDA / ROCm 等）提供了独立的 Simple Repository 索引，例如 /cpu/、/cu118/、/cu121/、/cu124/、/cu126/ 等。你可以根据使用的硬件与驱动版本选择对应的索引 URL。
使用方式
以 CUDA 12.6（cu126）为例：
pip
pip install torch torchvision torchaudio --index-url https://mirror.sjtu.edu.cn/pytorch-wheels/cu126/
或使用 --extra-index-url：
pip install torch torchvision torchaudio --extra-index-url https://mirror.sjtu.edu.cn/pytorch-wheels/cu126/
Astral uv
使用 uv add（推荐，锁定依赖）：
uv add torch torchvision torchaudio --index pytorch-cu126=https://mirror.sjtu.edu.cn/pytorch-wheels/cu126/
或使用 uv pip：
uv pip install torch torchvision torchaudio --index-url https://mirror.sjtu.edu.cn/pytorch-wheels/cu126/
在 pyproject.toml 中配置源（推荐，避免依赖混淆）：
[tool.uv.sources]
torch = { index = "pytorch-cu126" }
torchvision = { index = "pytorch-cu126" }
torchaudio = { index = "pytorch-cu126" }
[[tool.uv.index]]
name = "pytorch-cu126"
url = "https://mirror.sjtu.edu.cn/pytorch-wheels/cu126/"
explicit = true
Poetry
添加源并安装依赖：
poetry source add --priority=explicit pytorch-cu126 https://mirror.sjtu.edu.cn/pytorch-wheels/cu126/
poetry add --source pytorch-cu126 torch torchvision torchaudio
或者在 pyproject.toml 中配置：
[[tool.poetry.source]]
name = "pytorch-cu126"
url = "https://mirror.sjtu.edu.cn/pytorch-wheels/cu126/"
priority = "explicit"
[tool.poetry.dependencies]
torch = { version = "*", source = "pytorch-cu126" }
PDM
添加源并安装依赖：
pdm config pypi.extra.url "https://mirror.sjtu.edu.cn/pytorch-wheels/cu126/"
pdm add --source pytorch-cu126 torch torchvision torchaudio
或者在 pyproject.toml 中配置：
[[tool.pdm.source]]
name = "pytorch-cu126"
url = "https://mirror.sjtu.edu.cn/pytorch-wheels/cu126/"
type = "index"
以下基于 torch_stable.html 的 find-links 使用方式已废弃：
pip install torch torchvision torchaudio -f https://mirror.sjtu.edu.cn/pytorch-wheels/torch_stable.html