Installing PyTorch with ROCm on Radeon GPUs

PyTorch is the go-to framework for deep learning, and on AMD hardware it runs through ROCm. Once ROCm is installed, getting PyTorch working on a Radeon GPU is just a pip install away.

I tested this on a Radeon RX 6800M (gfx1031, RDNA2) with ROCm 7.14 on Ubuntu 26.04 — the exact commands below are what worked. The same flow applies to other RDNA2/RDNA3 cards; just swap the gfx target in the install command.

Prerequisites

ROCm installed and verified — if you haven't done this yet, follow our Installing ROCm on Ubuntu for Radeon GPUs guide first, and confirm rocminfo shows your GPU.
Python 3.11, 3.12, 3.13, or 3.14 installed.
Your GPU's gfx architecture (e.g. gfx1031 for RX 6800M).

Step 1 — Create a virtual environment

Always install PyTorch in a fresh virtual environment so it never conflicts with system packages or other Python projects:

python3.12 -m venv .venv

Use whichever Python version you have — python3.11, python3.13, or python3.14 all work. This creates a folder called .venv in your current directory.

Step 2 — Activate the environment

Activate it so python and pip point at the venv:

source .venv/bin/activate

You'll know it worked when the prompt shows (.venv) at the start.

Step 3 — Install PyTorch with ROCm support

Install the ROCm-enabled PyTorch, torchvision, and torchaudio from AMD's wheel repository. Use your GPU's device-gfx target — for the RX 6800M that's device-gfx1031:

python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ \
    "torch[device-gfx1031]==2.12.0+rocm7.14.0" \
    "torchvision[device-gfx1031]==0.27.0+rocm7.14.0" \
    "torchaudio==2.11.0+rocm7.14.0"

A few notes:

The [device-gfx1031] extra pulls in the ROCm libraries tuned for your exact GPU architecture — don't skip it.
For other cards, swap the target: gfx1030 (RX 6800/6700 XT), gfx1100 (RX 7900 XTX), and so on. Or use [device-all] if you want support for every architecture at once (larger download).
The +rocm7.14.0 version tag must match your installed ROCm release.

If pip complains about resolving dependencies, make sure the venv is active and you're passing --index-url https://repo.amd.com/rocm/whl-multi-arch/ — that repository is what AMD publishes the ROCm builds to.

Step 4 — Verify the GPU is detected

Run this one-liner to confirm PyTorch sees your AMD GPU:

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

It prints True if PyTorch and ROCm are installed correctly and your AMD GPU is detected. (Yes, it's cuda in the API — PyTorch keeps the same interface for ROCm so code is portable.)

You can go further and check the device name:

python -c "import torch; print(torch.cuda.get_device_name(0))"

On the RX 6800M this reports something like AMD Radeon RX 6800M.

What's next?

With PyTorch running on ROCm you can:

Train models and run inference with full GPU acceleration
Run LLMs locally — Hugging Face Transformers, text-generation-webui, and similar tools work out of the box
Use torchvision for image models and torchaudio for audio pipelines

The official guide has more detail if you need it: Install PyTorch for ROCm — AMD AI ecosystem docs.