This repo is a narrow-target Docker stack: Ollama v0.21.0 + ROCm 7.2.2 + gfx1151 (Strix Halo). "Helpful" generalizations have repeatedly broken it. Stay inside the rails below.
- Do not add Vulkan, CUDA, NVIDIA, multi-arch, or
HSA_OVERRIDE_GFX_VERSIONpaths. gfx1151 is native; alternatives are out of scope. - Do not modify files under
external/ollama/. It is a pinned submodule (Ollama v0.21.0). Upstream changes go through a submodule bump in a separate PR, not in-place edits. - Do not flip the rocBLAS prune in
docker/Dockerfilefrom "delete other arches" back to "keep only*gfx1151*". The latter pattern deletes 54 arch-agnostic fallback.datfiles and causes a page fault at first kernel call (seedocs/build-fixes.mdFix 2). - Do not drop or weaken layers in
scripts/validate.sh. Treatmake validate(andmake validate-fullfor the long-context layer) as the working contract. - Do not silently change Ollama, ROCm, or image-tag versions. They
appear in README,
docker-compose.yml,docker/Dockerfile, anddocs/; bump them together.
- Prefer editing existing files over creating new ones.
- When changing behaviour, update the matching prose in
docs/(validation-tests.md,build-fixes.md,break-modes.md,rocblas-prune.md). - Keep MES firmware logic in
scripts/install-mes-firmware.shidempotent and root-required; do not add silent network fetches outside the documented commit-pinned download. - Avoid hard-coded line-number references into
external/ollama/; link the file or symbol instead (lines drift on submodule bumps).
- Run
make mes-checkbefore debugging anything that looks like GPU init failure. - Run
./scripts/validate.sh --layer Nto isolate which ladder layer fails. - Read
docs/build-fixes.mdbefore proposing a "fix" for any page-fault, Vulkan-fallback, or CPU-fallback symptom; the answer is almost always already there.