fix(hipblas): symlink amdgpu.ids so ROCm backends find the ASIC ID table

[localai-bot]

Summary

• ROCm's bundled libdrm_amdgpu looks up the GPU ASIC ID table at a hardcoded fallback path, /opt/amdgpu/share/libdrm/amdgpu.ids, which only exists if AMD's full amdgpu-install (graphics/DKMS) stack is installed
• The hipblas image (rocm/dev-ubuntu-24.04 base) is compute-only, so that path never exists, and every model load on a hipblas backend logs /opt/amdgpu/share/libdrm/amdgpu.ids: No such file or directory
• Ubuntu's libdrm-common package ships the equivalent data at /usr/share/libdrm/amdgpu.ids; symlink it into place for BUILD_TYPE=hipblas, guarded so it's a no-op if either path is already in an unexpected state

Backends run as plain child processes of the main local-ai process and share its root filesystem (confirmed by tracing pkg/oci/image.go extraction + pkg/model/process.go/core/cli/worker exec), so a single symlink at the top-level Dockerfile (requirements-drivers stage) covers all hipblas backends (diffusers, llama.cpp, etc.) running under localai/localai:*-gpu-hipblas.

Fixes #10624

Test plan

• Pulled rocm/dev-ubuntu-24.04:7.2.1 directly and ran apt-get install hipblas-dev hipblaslt-dev rocblas-dev (the exact packages this Dockerfile installs for BUILD_TYPE=hipblas), then confirmed: /opt/amdgpu/share/libdrm/ does not exist (reproduces the reported bug) and /usr/share/libdrm/amdgpu.ids exists, owned by libdrm-common (dpkg -S amdgpu.ids)
• Verified the added RUN shell logic against a real docker build (mocked /usr/share/libdrm/amdgpu.ids present) — symlink is created correctly and resolves
• Inspected an existing published *-gpu-rocm-hipblas-llama-cpp backend image to confirm libdrm_amdgpu is not statically linked/bundled into the backend's own lib/, so its lookup relies on the outer image's filesystem, matching where this fix lives
• Full localai/localai:*-gpu-hipblas image build + on-hardware ROCm verification (needs AMD GPU, not available in this environment)

[gspeter-max]

hi @mudler
you should also apply this change to backend/Dockerfile.python.

The Python backends (such as vLLM and sglang) build hipblas variants and will face the same GPU identification warning or failure if the amdgpu.ids symlink is missing there.

[mudler]

hi @mudler you should also apply this change to backend/Dockerfile.python.

The Python backends (such as vLLM and sglang) build hipblas variants and will face the same GPU identification warning or failure if the amdgpu.ids symlink is missing there.

backends Dockerfiles are used to create containers from-scratch. these images are not executed as-is, so moving the files there has no effect. IF /opt/amdgpu/share/libdrm/amdgpu.ids can be moved and packaged to backends, that's another story. But requires readers of that file to have a way to specify a different path.