Running Ollama on AMD GPU (gfx10xx Series) — Setup Guide

This guide explains how to set up Ollama to use an AMD discrete GPU (like RX6700S, gfx1032 for G14 2022).

---

Steps

1. Find your GPU

• Open: %localappdata%\Ollama\server.log

• Look for the gpu_type field — you want gfx10xx (example: gfx1032 for G14 2022).

• Important:

• There are 2 GPUs — integrated and discrete. GPU 0 was the discrete GPU for me (you want to run it on your dGPU).

• Open Task Manager (Ctrl + Shift + Esc), go to the Performance tab and check the GPU type if you aren't sure.

---

2. Backup Current Install

• Go to: C:\Users%username%.ollama

• Make a copy of the .ollama folder and rename it with _backup at the end.

• (Optional but recommended): Create a Windows System Restore Point.

---

3. Download OllamaSetup.exe

• Download from: Ollama for AMD Releases
• Close Ollama if it’s running (system tray → Task Manager).
• Run the installer.
  (It’s unsigned — click "More Info" → "Run Anyway").
• After installation, check the hidden system tray to close the Ollama service that is running.

---

4. Install ROCm Libraries

• Go to: ROCmLibs Latest Release
• Download the latest version (e.g., 0.6.2.4 currently).
• Find your gfx version's zip file and download it.
• Some GPUs have multiple versions — read carefully before downloading.

If the link isn't working:

• Visit the Ollama for AMD Wiki.
• Navigate to the Demo Release Version section.
• Find the ROCm libraries for your version (e.g., rocm.gfx1035.for.hip.sdk.6.1.2.7z).

---

5. Replace ROCm Libraries

• Extract the .7z file.
  You should find:

  • A rocblas.dll file.
  • A library folder.

• Navigate to: %localappdata%\Programs\Ollama\lib\ollama

• Locate the current rocblas.dll:

• It may be in the main folder or inside a subfolder like rocm.

• Backup the original files (rename them with .old or _old if you wish).

• Copy:

  • rocblas.dll
  • library folder (usually inside rocm\rocblas\).

Folder structure may vary slightly from guide references.

---

6. Final Steps

• This is a brief overview — it's recommended to read the Wiki for any changes or updates.
• Important:
• Do not update Ollama normally via the app.
• Always manually re-download and reinstall
• Automatic updater is available byronleeee's releases.
• Issues?
• Open an issue on the likelovewant GitHub page.
• Restoring Models:
• If you backed up .ollama, you can restore previous models by copying back the models folder.
• Test GPU Usage:
• Start with a small model.
• Monitor GPU usage via Task Manager.
• Note: If the model is too large, some load might fall back to the CPU.

---

🔗 Links and References

• Ollama for AMD GitHub
• ROCm Libraries for AMD GPUs
• Ollama for AMD Wiki
• Troubleshooting AMD ROCm on Windows
• YouTube Walkthrough (Outdated)