STARBLOG / 阅读手记2026.01.05

Windows + AMD ROCm + PyTorch: A Debuff-Maxed Journey Training on a 6650 XT AMD GPU

Running PyTorch on Windows with a 6650 XT? AMD doesn't even officially support this card, so you have to compile ROCm yourself. After a night of wrestling with environment variables and compile errors, I didn't even get a single wheel out of it. So why insist on training on an AMD GPU on Windows?

曦远 (DealiAxy)2026.01.05约 30 分钟

目录17 个章节

Preface

Recently I've been seeing the z-image model a lot, and it seemed pretty interesting.

Embarrassingly enough, it's also been a while since I last tinkered with AI image generation. The servers at the studio have long been maxed out by large models, so piling on more tasks would be a bit inhumane.

So I set my sights on my own desktop.

The RX 6650 XT I bought a few years ago—back then I never even considered using it for training. But years later, AMD's ROCm has gradually matured.

It used to be something you could only tinker with on Linux, but this year ROCm v7 started supporting Windows, which suddenly made this somewhat feasible.

The only problem: PyTorch still only has official support on Linux, so on Windows you have to compile it yourself.

Alright then, let's give it a shot.

This article records my complete journey of pitfalls on the Windows 11 + AMD 6650 XT + ROCm + PyTorch path.

As of this article's publication, I still haven't succeeded 😂

But I'll treat this as an ongoing series, and keep writing until I succeed (or give up completely).

Virtual Environment

This step is basic Python stuff, so I'll go over it briefly.

Creating a Virtual Environment

mkdir d:/code/2/_temp
cd d:/code/2/_temp
python -m venv venv

The above commands will create a virtual environment at d:/code/2/_temp/venv, with the same Python version as the system install.

Of course, you can also use solutions like conda / mamba / uv.

I initially used Stability Matrix to install ComfyUI, which automatically created a virtual environment for me, so I just went ahead and used that.

Activating the Virtual Environment

Stay in the d:/code/2/_temp directory and use the following command

venv/scripts/activate

Then make sure to verify

pip --version

Make sure the output path points to the pip inside the virtual environment, not the system-wide one.

Installing ROCm

Next is the most critical step: installing ROCm, which you can think of as CUDA for AMD GPUs.

My GPU is an AMD RX 6650 XT, whose corresponding code name is gfx1032.

Confirming the GPU Architecture

You can look up the mapping between GPUs and architectures in the official ROCm docs:

https://rocm.docs.amd.com/en/latest/reference/gpu-arch-specs.html

There's another method: use the GPU-Z tool to check the GPU architecture. For example, this card of mine is Navi 23.

Then look up the corresponding code name in Mesa's source: https://gitlab.freedesktop.org/mesa/mesa/-/blob/main/src/amd/common/amd_family.c to find the matching code name in the source

About GPU Support

Note that AMD's official ROCm v7 does not support the 6x00/6x50 series GPUs.

That means for a card like the 6650 XT, you either use a third-party prebuilt version, or compile it yourself.

If your GPU is not on the official support list, you can visit this page to find the corresponding prebuilt packages: https://d2awnip2yjpvqn.cloudfront.net/v2/

If you happen to have an officially supported model (like the 7000 series), things are much simpler. Just use AMD's official packages. For example, here's the install for 7x00 series GPUs:

pip install --index-url https://rocm.nightlies.amd.com/v2/gfx110X-all/ "rocm[libraries,devel]"

And the install command for the 6650 XT is:

python -m pip install --upgrade --index-url https://d2awnip2yjpvqn.cloudfront.net/v2/gfx103X-dgpu/ rocm rocm-sdk-core rocm-sdk-devel rocm-sdk-libraries-gfx103x-dgpu

After installing, run a test

rocm-sdk test

If everything is fine, it will print a bunch of test results, ending with:

$ rocm-sdk test
...
Ran 22 tests in 258.284s
OK

Building PyTorch

After ROCm is installed, only one problem remains: PyTorch has no official Windows + ROCm release, so you have to compile it yourself.

Official docs: https://github.com/ROCm/TheRock/tree/main/external-builds/pytorch#build-instructions

Getting the Code

First clone the ROCm/TheRock repo, then enter the directory:

cd external-builds/pytorch

Use the official scripts to pull the PyTorch-related repos:

⚠️ Official reminder: on Windows, try to use shorter paths to avoid command length limits

python pytorch_torch_repo.py checkout --checkout-dir d:/code/2/pytorch
python pytorch_audio_repo.py checkout --checkout-dir d:/code/2/pytorch-audio
python pytorch_vision_repo.py checkout --checkout-dir d:/code/2/pytorch-vision

Only after pulling did I realize how huge these repos are. I'd recommend adding the --depth=1 flag so you don't pull the full git history.

Environment Setup

This is the official build command, which includes an option to install ROCm. I had already installed ROCm, so I didn't follow this command.

python build_prod_wheels.py build ^
  --install-rocm --index-url https://rocm.nightlies.amd.com/v2/gfx110X-all/ ^
  --pytorch-dir C:/b/pytorch ^
  --pytorch-audio-dir C:/b/audio ^
  --pytorch-vision-dir C:/b/vision ^
  --output-dir %HOME%/tmp/pyout

If you already have ROCm, you only need to run this command. But hold on—the prep work isn't done yet.

python build_prod_wheels.py build ^
  --pytorch-dir d:/code/2/pytorch ^
  --pytorch-audio-dir d:/code/2/pytorch-audio ^
  --pytorch-vision-dir d:/code/2/pytorch-vision ^
  --output-dir d:/code/2/_temp/pyout

Installing and Configuring the clang / MSVC Build Environment

You also need to install Visual Studio 2022's C++ development tools.

Download link: https://visualstudio.microsoft.com/thank-you-downloading-visual-studio/?sku=Community&channel=Release&version=VS2022&source=VSLandingPage&cid=2030&passive=false

The install takes a bit over 6 GB of space.

If you start compiling right after installing, it probably still won't work—it'll complain that it can't find VCToolsRedist or similar, because the paths haven't been added to the environment variables.

Adding Environment Variables

I first tried setting environment variables manually, for example:

# 1. 设置基础安装路径 (注意 VS 2022 默认通常在 'C:\Program Files', 确认你的在 x86)
$env:VCINSTALLDIR = "C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC"

# 2. 引用上面的变量来设置 Redist 路径 (使用双引号确保变量被解析)
$env:VCToolsRedistDir = "$($env:VCINSTALLDIR)\Redist\MSVC\14.44.35112\"

# 3. 验证是否设置成功
echo $env:VCToolsRedistDir

Although setting them manually can fix some errors, compiling PyTorch depends on dozens of complex environment variables (such as INCLUDE, LIB, PATH, etc.). Manually setting $env:VCToolsRedistDir only fixes this one error; later you might run into issues like cl.exe not being found.

Which I did, in fact 😂

Another approach: run the following commands in PowerShell to automatically load Visual Studio's full development environment (much more reliable than setting variables by hand):

# 导入 VS 编译环境到当前 PowerShell
Import-Module "C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\Common7\Tools\Microsoft.VisualStudio.DevShell.dll"
Enter-VsDevShell -VsInstallPath "C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools" -SkipAutomaticLocation -Arch amd64

The most reliable approach: search for and open "x64 Native Tools Command Prompt for VS 2022" from the Start menu (x64 Native Tools Command Prompt).

Starting the Build

Once the prep is done, you can finally start compiling:

python .\build_prod_wheels.py build --pytorch-dir d:/code/2/pytorch --pytorch-audio-dir d:/code/2/pytorch-audio --pytorch-vision-dir d:/code/2/pytorch-vision --output-dir d:/code/2/_temp

After the first compile started I went to sleep, with auto-shutdown set for when the build finished. When I got up the next day, there was nothing.

Started a second compile, and this time it errored out. I asked Grandpa Gemini the LLM to help analyze it, and it told me the clang-cl compiler had strict mode enabled, treating warnings as errors.

I tried turning that option off with environment variables:

PowerShell commands

# 允许缺少原型警告
$env:CFLAGS = "-Wno-error=missing-prototypes -Wno-missing-prototypes -Wno-error"
$env:CXXFLAGS = "-Wno-error=missing-prototypes -Wno-missing-prototypes -Wno-error"

cmd commands (note: no spaces on either side of the equals sign)

set CFLAGS="-Wno-error=missing-prototypes -Wno-missing-prototypes -Wno-error"
set CXXFLAGS="-Wno-error=missing-prototypes -Wno-missing-prototypes -Wno-error"

# 验证一下
echo %CFLAGS%, %CXXFLAGS%

But it still errored after I changed that, so I tried clearing the build cache, because RegisterMeta_0.cpp in the error message is a generated intermediate file, and leftover cache can sometimes mess up the build logic.

Deleted the d:/code/2/pytorch/build folder and started the build again. This time no errors for now, and the CPU was maxed out at 100%.

Error Analysis

It still errored out in the end

Error Messages

The error was too long

I'll paste the main errors here:

error: use of undeclared identifier 'Wno'
error: use of undeclared identifier 'error'
error: use of undeclared identifier 'missing'
error: use of undeclared identifier 'prototypes'

The corresponding context:

static const std::map<string, string> kMap = CAFFE2_BUILD_STRINGS;

And CAFFE2_BUILD_STRINGS was expanded like this in the auto-generated header file:

{"CXX_FLAGS", ""-Wno-error=missing-prototypes -Wno-missing-prototypes" /DWIN32 /D_WINDOWS ..."}

Note this double-quote structure:

""-Wno-error=missing-prototypes -Wno-missing-prototypes"

This is illegal C/C++ string concatenation, which directly causes the macro-expanded code to not be valid C++ syntax.

So it was probably the CFLAGS and CXXFLAGS environment variables I set earlier that were wrong?

Trying a Downgrade

This time Grandpa Gemini the LLM wasn't of any use, so I had to turn to Grandpa GPT, who told me it might be that the PyTorch source version was too new (I was using 2.11 here).

I decided to try downgrading to 2.10. Using the method shared in Building ROCm 7.1 and PyTorch on Windows for Unsupported GPUs: My Hands-On Guide, you can specify the version:

python pytorch_torch_repo.py checkout --repo-hashtag release/2.10 --gitrepo-origin https://github.com/ROCm/pytorch.git --depth=1 --checkout-dir d:/code/2/pytorch
python pytorch_audio_repo.py checkout --require-related-commit --depth=1 --checkout-dir d:/code/2/pytorch-audio --torch-dir d:/code/2/pytorch
python pytorch_vision_repo.py checkout --require-related-commit --depth=1 --checkout-dir d:/code/2/pytorch-vision --torch-dir d:/code/2/pytorch

Damn, it still didn't work. This time GPT doubled down on me, insisting that you can't compile PyTorch on Windows...

Fine, I'll jot this down for now.

References

_EOF_