2 Solution logo2 SOLUTION
ITENESFRDE
Research · Artificial intelligence

The first
in the world

PyTorch with ROCm, native on Windows, on the Radeon VII: a combination that nobody, as far as public sources show, had ever made work.

2 Solution · November 2026 · 4-minute read

The Radeon VII is a 2019 AMD card with the “Vega” architecture (gfx906): 16 GB of very fast memory and compute power that is still respectable today. But for artificial intelligence it has become an orphan card: official support has ended, and on Windows scientific-computing software never really considered it in the first place.

We put it back to work. Not with a compatibility trick, but by rebuilding from scratch the entire software chain, from the compiler up to the library used by researchers all over the world: PyTorch.

What we did first

01
PyTorch compiled natively with ROCm on Windows for gfx906 Project “LA RUOTA”: no intermediate translation layer, about 5 times faster than the previous route via ZLUDA.
02
All the cuDNN operations used by PyTorch translated for AMD GPUs 86 symbols out of 86 implemented and verified in ZLUDA towards MIOpen: convolutions, normalizations, recurrent networks.
03
MIOpen brought back to working order on the Radeon VII with current drivers Kernels recompiled for the format accepted by the driver and three silent defects fixed.
04
Training modern language models on a 2019 card LoRA fine-tuning of a 4-billion-parameter model, raised from 182 to 562 tokens per second.

Why it matters

Along the way we found a defect that no error message reported: for one specific combination of shapes, the library picked the fastest computation method, but not the correct one. The result was a wrong gradient precisely in the first layer of every network that works on color images. Training did not stop: it simply did not learn.

Automatic search picks the fastest method, not the right one. Every new shape must be checked at least once against a reference computation.

It is the lesson we bring to every project, not only in artificial intelligence: a result is not verified until it has been compared with reality. On this basis we measured everything: 98 operations out of 98 correct, 20 out of 20 in GPU linear algebra, hundreds of convolution shapes checked one by one.

What it means for our clients

It means knowing how to run artificial intelligence on the hardware you already have, on site, without sending confidential data to the cloud. And it means a method: when something is not supported, we do not stop at “it can't be done”, but go down a level until we find the cause.

We share part of this work as free software, so that others can bring their own cards back to life.

The repositories on GitHubAI for your company

“The first in the world”: as far as the public sources consulted (official documentation, repositories and reference forums) show as of November 2026, there are no other documented cases of PyTorch compiled natively with ROCm on Windows for gfx906 GPUs. If you know of one, write to us: we will gladly cite it. ROCm, MIOpen, PyTorch and ZLUDA are projects of their respective authors, distributed under their own licenses.