dropping integrated GPU; to enable, set OLLAMA_IGPU_ENABLE=1

What you see

A fresh Ollama install on a Strix Halo box. Filter the service journal for GPU lines and two entries come out that read as a contradiction:

msg="dropping integrated GPU; to enable, set OLLAMA_IGPU_ENABLE=1"
     id=0 library=Vulkan name=Vulkan0 description="AMD Radeon Graphics (RADV GFX1151)"
msg="inference compute" id=0 library=ROCm compute=gfx1151
     description="Radeon 8060S Graphics" type=iGPU total="117.4 GiB" available="117.4 GiB"

The first line reads as “your GPU is unsupported”. People take it at face value and head for the issue tracker.

The second observable is quieter. A model loads and answers, but slowly, and ollama ps shows a PROCESSOR value like XX%/XX% CPU/GPU instead of 100% GPU. Nothing in the chat output tells you.

Where we saw it

• Hardware: Strix Halo, Ryzen AI Max+ 395, one of our own boxes. The iGPU appears as Radeon 8060S Graphics under ROCm and as AMD Radeon Graphics (RADV GFX1151) under Vulkan.
• Runtime: Ollama, fresh default install from the official script (curl -fsSL https://ollama.com/install.sh | sh). The guide pins no version; it calls it “a current Ollama”.
• Date: 2026-09-01, the guide’s date; the guide states it was written while the install was being performed.

Cause

Ollama ships several backends in one bundle. Its Vulkan path drops integrated GPUs by default, and that is all the first line says. The bundled ROCm backend recognizes gfx1151 directly, takes the device, and reports the full memory ceiling. Both lines are true at the same time; the Vulkan drop is not a verdict on the hardware.

OLLAMA_IGPU_ENABLE=1 only matters if you specifically want the Vulkan path. On a current Ollama the device works through ROCm without it.

The CPU fallback has two causes the guide names: the model is too big for the memory ceiling set by the TTM kernel parameter, or the GPU was not picked up at all. Either way answers still arrive, with part or all of the model running on the CPU.

Fix

Nothing to fix in the drop line itself. Read the journal, look for the ROCm line, then prove where the model landed:

journalctl -u ollama -n 80 | grep -iE "gpu|rocm|gfx|compute"

Drop line plus a library=ROCm compute=gfx1151 line: leave OLLAMA_IGPU_ENABLE alone and move on.

Drop line and no ROCm line: your Ollama predates the bundled gfx1151 support. Update it before touching anything else.

Then, with a model loaded:

ollama ps

PROCESSOR must say 100% GPU. Ours does. If it reads XX%/XX% CPU/GPU, either the model exceeds the ceiling from the memory allocation page, or the GPU was not detected: go back to the journal.

Still open upstream?

Not tracked by us. The guide links no upstream issue or PR.

Evidence

• Ollama on Strix Halo: install, GPU detection, first model
• Strix Halo memory allocation, the ceiling the fallback check depends on