HSA_OVERRIDE_GFX_VERSION: ROCm on unlisted Radeons

What you will see

HSA_OVERRIDE_GFX_VERSION

What it means

An environment variable that makes ROCm treat your Radeon as a nearby supported model, so ROCm builds of Ollama and llama.cpp can use a card ROCm does not list. Linux only.

Why it happens

• You need it when the card's gfx target is missing from ROCm's list — the symptom is a rocBLAS "Cannot read TensileLibrary" error, or Ollama falling back to the CPU.

How to fix it

1. Find your target with rocminfo | grep gfx, then set the version of the nearest supported one in x.y.z form: 10.3.0 stands for gfx1030, 11.0.0 for gfx1100.
2. Set it for the server, not your shell: for the Ollama service, sudo systemctl edit ollama and add Environment="HSA_OVERRIDE_GFX_VERSION=10.3.0", then restart it.
3. With two different Radeons, set each one: HSA_OVERRIDE_GFX_VERSION_0=10.3.0, HSA_OVERRIDE_GFX_VERSION_1=11.0.0.
4. On Windows it does not apply: Ollama there uses ROCm on RX 7600–7900 and Vulkan on other Radeons.

Guides that cover this

• AMD GPU + llama.cpp via ROCm (Quick Start)
• Ollama on Windows (Native, No WSL)

Related errors

• rocBLAS error: Cannot read TensileLibrary.dat
• Ollama shows 100% CPU instead of the GPU