A ROCm 10 distribution built from ROCm/TheRock targeting gfx906 specifically — a GPU family that ROCm 10 doesn't officially support upstream (official support starts at gfx908). Built from source via the Wizard815/TheRock-gfx906 fork, which carries the one patch needed to get a gfx906-scoped build working (a subproject step that otherwise assumes a newer architecture is present) plus the full build recipe.
:sdk — the full build output: headers, hipcc, static libs, CMake package configs. Use this as a FROM base when you need to compile something against ROCm 10 (e.g. building llama.cpp, PyTorch, or anything else with HIP support) — there are no prebuilt wheels for this ROCm-version-on-gfx906 combination, everything has to build from source.:runtime — just the shared libraries (rocBLAS, MIOpen, the HIP runtime, etc.), no compiler or headers. Use this as the final stage once you already have a compiled binary — smaller image, smaller attack surface.zenth815/mx-llama-rocm10-gfx906 — mx-llama.cpp built against this image, with gfx906-specific optimizations ported in from the community.
Content type
Image
Digest
sha256:af683a1e7…
Size
2.7 GB
Last updated
about 1 month ago
docker pull zenth815/rocm10-gfx906:runtime