chore(rocm): ROCm 7.14 / TheRock packaging support for AMD builds (base bump + build compat)

[walcz-de]

What

Brings LocalAI's ROCm / hipblas build up to ROCm 7.14 and makes it compatible with AMD's new "TheRock" packaging.

This is a platform-wide change for AMD, not a single-GPU one. The base image, package layout and library paths it fixes apply to every architecture in AMDGPU_TARGETS (gfx908 / gfx90a / gfx942 / gfx1030 / gfx1100 / gfx1101 / gfx1102 / gfx1151 / gfx1200 / gfx1201) — the ROCm/hipblas images simply do not build (and then do not load) on ROCm 7.14 until these are fixed, regardless of the target card. gfx1151 / Strix Halo is only the AMD hardware I had available to runtime-validate on.

• Bumps the ROCm base rocm/dev-ubuntu-24.04:7.2.1 → :7.14.0-full across backend-matrix.yml, base-images.yml, image.yml/image-pr.yml, the .agents/* docs and the Dockerfile.base-grpc-builder comment.
• Fixes the ROCm 7.x / TheRock build breakages (below).

Follow-up to #10883, companion to the docs PR #9229.

Why — ROCm 7.x switched to "TheRock" packaging

See AMD's TheRock transition guide. A naive base bump breaks the AMD build in three ways (all fixed here, all architecture-agnostic):

1. Legacy -dev metapackages removed. apt-get install hipblas-dev hipblaslt-dev rocblas-dev fails (Unable to locate package) — they are consolidated + arch-split (amdrocm-blas<ver>-gfx*, amdrocm-blas-dev, …). The -full base already ships them, so the install is dropped in both places that used the legacy names (Dockerfile requirements-drivers + .docker/install-base-deps.sh). The amd capability marker + ldconfig are kept.

2. ROCm libs are scattered and not on the linker path. TheRock places libs under /opt/rocm/lib, /opt/rocm/lib/rocm_sysdeps/lib and /opt/rocm/llvm/lib with no ld.so.conf.d entry. The ldd-based GPU-lib packaging (scripts/build/package-gpu-libs.sh) then silently skips librocm_kpack, librocm_sysdeps_*, rocBLAS, hipBLASLt, etc., and the built backend fails to load at runtime on the GPU (... cannot open shared object file). CI does not catch this — it only surfaces when the backend loads on an AMD GPU. Fixed by registering the three ROCm lib dirs in /etc/ld.so.conf.d/rocm.conf before ldconfig, in both blocks.

3. The base bump itself. base-images.yml rebuilds the base-grpc-rocm-amd64 gRPC cache on 7.14 so it stays ABI-compatible with the new runtime base.

Testing

The build changes are architecture-agnostic — they fix the ROCm/hipblas image for all AMD architectures. I runtime-validated on the AMD hardware I have, gfx1151 / Strix Halo (Ryzen AI Max+ 395 / Radeon 8060S): the core image and the llama-cpp backend build, all ROCm libraries bundle (verified ldd resolves everything), and inference runs on the GPU — model loaded, tokens generated, no missing-lib errors.

Since I only have one AMD GPU to test on, a maintainer or CI with other AMD cards (CDNA gfx908/90a/942, RDNA gfx1030/1100/…) should confirm — especially fix #2, which is a runtime-only failure.

(Aside, out of scope: GGML_HIP_ROCWMMA_FATTN is not enabled in the llama-cpp build flags — a potential RDNA3 flash-attention perf win.)

🤖 Generated with Claude Code

https://claude.ai/code/session_01NK35FML8epTCLBGwCor9et

[localai-org-maint-bot]

I reviewed this and found one missed ROCm 7.14 build path: backend/Dockerfile.golang still installed the removed hipblas-dev, hipblaslt-dev, and rocblas-dev packages. I pushed the matching TheRock linker-path migration as 8c08804a.

The content checks pass locally (git diff --check, shell syntax, no remaining obsolete ROCm package installs, and no stale 7.2.1 build pins), but DCO correctly rejects my bot-authored commit. @walcz-de, please amend/recreate that commit under your identity with your own Signed-off-by certification and force-push the branch. Once DCO and the substantive CI checks are green, I can finish the review and flag @mudler.

[walcz-de]

Thanks for the catch — backend/Dockerfile.golang was indeed a fourth spot with the removed legacy metapackages. I've re-authored your fix under my identity with my own Signed-off-by (kept you as Co-authored-by) and force-pushed (23b2a98c), so DCO should be green now.

That makes it four ROCm build paths migrated to TheRock: the core Dockerfile, .docker/install-base-deps.sh, backend/Dockerfile.python, and now backend/Dockerfile.golang.

[localai-org-maint-bot]

The contributor-owned re-authored fix is correct as far as matching the other three paths, and DCO is now green. However, the current head has one substantive ROCm regression that blocks approval:\n\nThe hipblas image reaches CMake and fails at because neither nor exists in the image after this PR removes the dev-package installation. This is job 88593150531, not infra noise. The assumption in the new comments that the image alone supplies all BLAS development files is therefore false for at least this backend.\n\nAMD's TheRock packaging docs map hipBLAS/rocBLAS/hipBLASLt into , and the component install guide says Debian/Ubuntu development files come from . Please install that package in the hipblas setup paths (at minimum where CMake-based Go/C++ backends need it; keeping the shared paths consistent is preferable), then confirm the qwen3-tts hipblas build reaches compilation.\n\nThe two extras-backend failures are unrelated to this PR: Transformers is the known sync-servicer/async-interceptor failure, and Moonshine is outside the changed ROCm build path. The cancelled matrix siblings are fail-fast fallout from the real qwen3 failure.\n\nOfficial package references:\n- https://rocm.docs.amd.com/en/latest/about/transition-guide-TheRock.html\n- https://rocm.docs.amd.com/projects/hipBLASLt/en/develop/install/install.html\n\n@mudler this is close, but the qwen3-tts ROCm build is a real blocker on the current head.

[localai-org-maint-bot]

The contributor-owned re-authored fix correctly makes backend/Dockerfile.golang match the other three paths, and DCO is now green. However, current head 23b2a98 has one substantive ROCm regression that blocks approval.

The qwen3-tts-cpp hipblas image reaches CMake and fails at find_package(hipblas) because hipblasConfig.cmake and hipblas-config.cmake are absent from rocm/dev-ubuntu-24.04:7.14.0-full after this PR removes the development-package installation. This is job 88593150531, not infrastructure noise. The new comment saying the -full image alone supplies all BLAS development files is therefore false for at least this backend.

AMD maps hipBLAS, rocBLAS, and hipBLASLt into amdrocm-blas under TheRock, and its component install guide says Debian and Ubuntu development files come from amdrocm-blas-dev. Please install that development package in the hipblas setup paths, at minimum where CMake-based Go and C++ backends require it, then confirm the qwen3-tts hipblas build reaches compilation.

The two extras-backend failures are unrelated: Transformers is the known sync-servicer and async-interceptor object Result cannot be used in await expression failure, and Moonshine is outside this ROCm build path. Cancelled matrix siblings are fail-fast fallout from the real qwen3 failure.

Official references:
https://rocm.docs.amd.com/en/latest/about/transition-guide-TheRock.html
https://rocm.docs.amd.com/projects/hipBLASLt/en/develop/install/install.html

@mudler this is close, but the qwen3-tts ROCm build is a real blocker on the current head.

Dismissed duplicate review submitted with stripped Markdown due to a client-side shell quoting error; the complete review at 2026-07-23T20:04:18Z supersedes it.

[walcz-de]

Confirmed — this is a real regression, not infra noise, and the diagnosis is precise. Thanks for pinning it to job 88593150531 and find_package(hipblas).

The distinction this PR got wrong: rocm/dev-ubuntu-24.04:7.14.0-full ships the runtime ROCm libraries, but not the CMake package-config files (hipblasConfig.cmake / hipblas-config.cmake) that find_package() resolves against. So dropping the legacy -dev metapackages is correct for the pure-runtime paths (where hipblas-dev/hipblaslt-dev/rocblas-dev now 404), but wrong for any backend that compiles hipBLAS-linked code from source — those need the TheRock development component.

Under TheRock, hipBLAS/rocBLAS/hipBLASLt consolidate into amdrocm-blas, and the Debian/Ubuntu development files come from amdrocm-blas-dev (matching your reading of the component install guide). The fix is therefore not "re-add the old metapackages" but "install amdrocm-blas-dev in the from-source CMake backend paths" — at minimum the hipblas Go/C++ backend setup (backend/Dockerfile.golang and the qwen3-tts hipblas build) — while leaving the runtime-only paths dev-free.

I'll push that (scoped to the CMake-consuming backends, not a blanket re-add) and re-validate on gfx1151 / Strix-Halo hardware with a real runtime model load — the same methodology that surfaced the CI-invisible ld.so.conf packaging gap (Fix #3), which a green build alone doesn't catch. I'll confirm the qwen3-tts hipblas build reaches compilation before flagging for re-review.

Agreed the Transformers/Moonshine failures are unrelated (known sync-servicer / out-of-path).

[localai-org-maint-bot]

The ROCm blocker is technically resolved at d569a90: amdrocm-blas-dev is now installed in the from-source Go/C++ hipBLAS builder, exactly where find_package(hipblas) needs the CMake package configuration, while runtime-only paths remain unchanged. The incremental diff is clean and DCO passes.

One repository-policy blocker remains before approval. Commits 23b2a98 and d569a90 both identify localai-org-maint-bot with a Co-authored-by trailer. LocalAI explicitly prohibits AI Co-authored-by trailers; AI involvement must use the Assisted-by format documented in .agents/ai-coding-assistants.md. Please reword both commits to remove those Co-authored-by trailers and replace them with an Assisted-by trailer such as:

Assisted-by: Codex:gpt-5

Keep your own Signed-off-by trailer intact when rewording. Once the head advances with policy-compliant attribution and DCO remains green, the technical change is ready for approval. @mudler the actual ROCm build issue is addressed; attribution is the remaining blocker.

[walcz-de]

Reworded both commits (7137cdb, 65ff127) to comply with the AI-attribution policy in .agents/ai-coding-assistants.md: the Co-authored-by: localai-org-maint-bot trailers are removed and replaced with Assisted-by trailers (Assisted-by: Codex:gpt-5 for the maintenance-bot-authored Go Dockerfile change, Assisted-by: Claude:claude-opus-4-8 for the amdrocm-blas-dev fix). My own Signed-off-by is intact on both and the tree is byte-for-byte identical to the previous head — message-only reword. DCO should stay green; ready for re-review.

[localai-org-maint-bot]

Re-review complete at 65ff127. The two rewritten commits now comply with .agents/ai-coding-assistants.md: the prohibited AI Co-authored-by trailers are gone, the appropriate Assisted-by trailers are present, and the contributor-owned Signed-off-by trailers remain intact. I also confirmed the resulting tree is byte-for-byte identical (tree 6e938e16) to the previously reviewed technical fix at d569a90. Fresh git diff --check, changed-shell syntax checks, DCO, and mergeability all pass. The ROCm 7.14 / TheRock change is ready from my side. @mudler