fix(gpu-libs): bundle transitive deps of GPU runtime libs (#10537)

[localai-bot]

What

Fixes #10537 — the ROCm llama-cpp backend (and every other GPU backend) fails to load with:

librocprofiler-register.so.0: cannot open shared object file: No such file or directory

Root cause

scripts/build/package-gpu-libs.sh is the shared GPU-lib packager used by all GPU backends. Its per-vendor packagers (package_rocm_libs, package_cuda_libs, package_intel_libs) copy a fixed allowlist of top-level runtime libs (libamdhip64, libhipblas, librocblas, ...) but never resolved their transitive dependencies.

Backends run through the bundled lib/ld.so with LD_LIBRARY_PATH=lib (see run.sh), so the binary only sees what's in lib/. On recent ROCm (base image rocm 7.2.1) the runtime libs link against librocprofiler-register.so.0, which is not in the allowlist, so it's never bundled — a hard load failure. libhipblas.so itself loads fine (it is allowlisted), which matches the issue report.

This is not related to the llama.cpp shared-libs change.

The Vulkan path already solved this class of problem with copy_elf_deps (ldd-based transitive resolution), but that sweep was only wired into the Vulkan ICD path.

Fix

• Add a generic sweep_transitive_deps that runs the existing copy_elf_deps (ldd) over everything the allowlist already bundled. ldd returns the full recursive closure, so one pass suffices.
• Wire it into the ROCm, CUDA and Intel packagers (Vulkan already had it).
• Core libc-family deps are skipped via is_core_lib, so we never shadow the loader's own libc/libstdc++.

Because the script is shared, this fixes the missing-transitive-dep class for llama-cpp, whisper, stablediffusion-ggml, turboquant, ds4, and the rest in one place.

Testing

• Added scripts/build/package-gpu-libs_test.sh: a self-contained regression test (gcc + ldd) that fabricates a primary lib linking a transitive lib, simulates the allowlist step, and asserts the sweep bundles the transitive dep. Red before the fix, green after.
• Smoke-tested the cpu and hipblas-without-/opt/rocm paths (graceful no-op / warning).

Note: the mechanism is verified locally with synthetic libs; full end-to-end validation runs in the ROCm build image in CI, using the same ldd resolution the Vulkan path already uses in production.

🤖 Generated with Claude Code

[localai-bot]

Validation against the real released image

Reproduced and verified against quay.io/go-skynet/local-ai-backends:latest-gpu-rocm-hipblas-llama-cpp (the image a v4.5.2 user runs).

Bug reproduced: the image bundles libhipblas.so.3, librocblas.so.5, libamdhip64.so.7 but librocprofiler-register.so.0 is absent from the entire image.

Confirmed it's a NEEDED dep, not dlopen (so ldd catches it and the sweep works):

$ readelf -d lib/libamdhip64.so.7
 (NEEDED)  Shared library: [librocprofiler-register.so.0]
 (NEEDED)  Shared library: [libhsa-runtime64.so.1]

Full audit of every bundled lib's NEEDED entries vs what's actually in lib/ (excluding core libc-family) shows librocprofiler-register is just the first of 8 missing transitive deps:

MISSING: libdrm_amdgpu.so.1
MISSING: libdrm.so.2
MISSING: libelf.so.1
MISSING: libnuma.so.1
MISSING: librocprofiler-register.so.0
MISSING: librocroller.so.1
MISSING: libz.so.1
MISSING: libzstd.so.1

A naive "add librocprofiler-register to the allowlist" fix would just move the failure to the next missing lib. The generic sweep_transitive_deps in this PR catches all 8 in one ldd pass (none are in the is_core_lib skip list), and the same applies to the CUDA/Intel packagers.

Residual note: this was validated by static readelf/ldd analysis of the released artifact. The sweep copies these from /opt/rocm + system libs at build time; final end-to-end confirmation happens when the ROCm backend is rebuilt in CI with this change.