AMD YES!

尽管AMD并没有在Windows上为RDNA2提供官方的ROCm支持，但是如果我们稍加搜索，依然可以找到可用的nightly版本。

环境准备

驱动

正常安装 AMD Software: Adrenalin Edition

uv

在终端中执行：

irm https://astral.sh/uv/install.ps1 | iex

Python环境

首先，编写pyproject.toml

[project]
name = "foo"
version = "0.1.0"
requires-python = ">=3.11,<3.12"

dependencies = [
    "torch==2.13.0a0+rocm7.14.0a20260612",
    "transformers>=5.6.0",

    "rocm==7.14.0a20260612",
    "rocm-sdk-core==7.14.0a20260612",
    "rocm-sdk-devel==7.14.0a20260612",
    "rocm-sdk-libraries-gfx103X-all==7.14.0a20260612",
]

[[tool.uv.index]]
name = "amd"
url = "https://rocm.nightlies.amd.com/v2-staging/gfx103X-all/"
explicit = true

[[tool.uv.index]]
name = "tencent"
url = "http://mirrors.cloud.tencent.com/pypi/simple"
default = true

[tool.uv.sources]
torch = { index = "amd" }
rocm = { index = "amd" }
rocm-sdk-core = { index = "amd" }
rocm-sdk-devel = { index = "amd" }
rocm-sdk-libraries-gfx103X-all = { index = "amd" }

[tool.uv]
package = false

随后，把wheel包安装到本地，在终端中执行：

uv sync

验证

现在你已经有了一个ROCm+pytorch环境，试试：

>>> import torch
>>> print(torch.__version__, torch.cuda.get_device_name(0))
2.13.0a0+rocm7.14.0a20260612 AMD Radeon RX 6800S
>>> a = torch.Tensor([1]).cuda()
>>> print(a+a)
tensor([2.], device='cuda:0')