6b294657ac
The current ollama:rocm image bundles ROCm 7.2, which natively supports the RDNA 3/3.5 mobile parts (780M gfx1103, 890M gfx1150, Strix Halo gfx1151). Forcing HSA_OVERRIDE_GFX_VERSION=11.0.0 coerces them onto gfx1100's kernels: unnecessary on the 890M and a source of faults on the 780M, which lacks the gfx1100 WMMA instructions those kernels use. - Map gfx1103/gfx1150/gfx1151 to no override (was 11.0.0) so ROCm discovers them natively. - Change the no-marker default from 11.0.0 to null. A hardcoded default gets more wrong as ROCm adds native targets; native discovery is the safer forward-looking default. Hardware that still needs coercion (e.g. RDNA 2 iGPUs on an install without the gfx marker) can force a value via the ai.amdHsaOverride KV. - Keep gfx1031..gfx1036 (RDNA 2 iGPUs like the 680M) on 10.3.0 — still not natively supported. Verified on NOMAD2 (Radeon 890M): after a Force Reinstall the container has no HSA_OVERRIDE, Ollama discovers the GPU as its native gfx1150 with the rocm_v7_2 libraries, and a model loads at 100% GPU with no faults. Depends on #1056 context; see companion issue for the 780M follow-up. Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> |
||
|---|---|---|
| .. | ||
| controllers | ||
| exceptions | ||
| jobs | ||
| middleware | ||
| models | ||
| services | ||
| utils | ||
| validators | ||