Will she run on your machine?

Windows GPU and AMD requirements

Implementation reviewed October 31, 2026. These are source-based requirements, not GPU benchmark results.

Does Local Waifu support AMD GPUs on Windows?

The Windows app detects AMD and, when the driver exposes ROCm v7/HIP7, downloads a matching Ollama ROCm runtime. Ollama's Vulkan runner is also bundled and is the supported fallback for Radeon cards that do not expose ROCm v7 on Windows (notably some RX 6000 systems). Supported ROCm Windows cards currently include RX 7900 XTX/XT/GRE, 7800 XT, 7700 XT, 7600 XT and 7600, plus Radeon PRO W7900/W7800/W7700/W7600/W7500; other cards may still accelerate through Vulkan. On a laptop with both iGPU and dGPU, Settings → Hardware has an optional Vulkan device field for `GGML_VK_VISIBLE_DEVICES`. Acceleration depends on the exact card, driver and runtime, so not every AMD GPU is guaranteed to work. Initial setup needs internet. If GPU setup fails, local chat can fall back to the bundled CPU runner, which can be slower.

Does AMD VRAM decide the automatic model tier?

Yes on current Windows builds: the app reads AMD dedicated VRAM from the display-adapter registry and uses the higher of the system-RAM and VRAM tiers. Settings → Hardware reports whether the active runtime has a ROCm runner and whether Ollama's Vulkan fallback is present. The thresholds stay conservative — a 16 GB card alone still lands in Light, while 24 GB can reach Balanced. AMD runtime provisioning and tier selection are separate mechanisms, and a recommendation never guarantees full GPU placement.

What are the actual model-tier thresholds?

Light extends through 18 GB; Balanced is 19-34 GB; Premium is 35-56 GB; Power begins at 57 GB. The selector takes the higher of the system-RAM tier and detected-VRAM tier, not their sum. A 16 GB GPU alone remains Light. The supported minimum is still 8 GB system RAM. A recommendation does not guarantee full GPU placement or a particular reply speed.

What if the model runs on CPU despite having a GPU?

Check Settings > Hardware and use Retry GPU setup when offered. The app can download a ROCm runtime and restart the engine; if the card does not expose ROCm v7, verify that Vulkan is available and check acceleration after loading a model. If that does not help, restart the app. Check the exact card against Ollama's hardware support, noting that current upstream support does not prove support in the app's pinned runtime.

Read the Windows GPU and tier guide for the inspected selection logic and the hardware settings guide for troubleshooting.