ci: refactor llama-cpp variant Dockerfiles to consume prebuilt base-grpc images (PR 2/2)

[localai-bot]

Summary

PR 2 of 2. Refactors Dockerfile.llama-cpp, Dockerfile.ik-llama-cpp, Dockerfile.turboquant to FROM the prebuilt base-grpc-* images shipped in PR #9737. Drops ~250 lines from each variant Dockerfile (gRPC stage + apt deps + CUDA/ROCm/Vulkan toolchain installs all live in the base now). On a LLAMA_VERSION bump, the gRPC compile (~25–35 min cold) is skipped — the variant build only re-compiles llama.cpp itself.

Commits

• A e61ddedf — backend_build.yml adds optional builder-base-image input + BUILDER_BASE_IMAGE build-arg. Backward-compatible.
• B f2867878 — refactor 3 variant Dockerfiles to FROM ${BUILDER_BASE_IMAGE}. Each shrinks ~300 → ~70 lines.
• C 6142f986 — backend.yml + backend_pr.yml forward builder-base-image from matrix.
• D 5a2e7115 — matrix entries: 21 of 23 llama-cpp-derived entries gain builder-base-image.
• E 79c7c7b3 — adds the missing base-grpc-l4t-cuda-12-arm64 variant to base-images.yml matrix and maps the 2 legacy JetPack entries (-nvidia-l4t-arm64-llama-cpp, -nvidia-l4t-arm64-turboquant) to it.

After all 5 commits: 23 of 23 llama-cpp-derived matrix entries are mapped.

What changes

Variant Dockerfile	Before	After
Dockerfile.llama-cpp	320 lines (gRPC stage + apt + CUDA/ROCm/Vulkan installs + compile)	80 lines (FROM prebuilt + compile)
Dockerfile.ik-llama-cpp	305 lines	69 lines
Dockerfile.turboquant	317 lines	77 lines

The compile step + ccache mount + package step are unchanged.

Mapping

(build-type, platforms)	builder-base-image
('', linux/amd64)	:base-grpc-amd64
('', linux/arm64)	:base-grpc-arm64
('cublas' cuda 12, linux/amd64)	:base-grpc-cuda-12-amd64
('cublas' cuda 13, linux/amd64)	:base-grpc-cuda-13-amd64
('cublas' cuda 13, linux/arm64)	:base-grpc-cuda-13-arm64
('hipblas', linux/amd64)	:base-grpc-rocm-amd64
('vulkan', linux/amd64)	:base-grpc-vulkan-amd64
('vulkan', linux/arm64)	:base-grpc-vulkan-arm64
('sycl_*', linux/amd64)	:base-grpc-intel-amd64
('cublas' cuda 12 + JetPack base, linux/arm64)	:base-grpc-l4t-cuda-12-arm64 (new)

Bootstrap before merging

PR 2's matrix references base-grpc-l4t-cuda-12-arm64 which doesn't exist on quay yet (PR 1 only built 9 variants; this PR adds the 10th). The branch's workflow_dispatch of base-images.yml has been triggered — see run 25610460480.

The 9 existing variants will hit BuildKit cache and finish in ~5 min; the new l4t-cuda-12-arm64 variant builds cold (~30 min). Merge after that run completes successfully.

Test plan

• Bootstrap dispatch run on the branch completes — all 10 base-grpc-* tags are pushed to quay.
• Merge PR.
• Next master push that touches backend/cpp/llama-cpp/ (e.g. an organic LLAMA_VERSION bump) schedules the per-arch -cpu-llama-cpp build, the merge job, and the legacy -nvidia-l4t-arm64-llama-cpp build. All consume their respective prebuilt bases. Expected wall-clock: ~25–35 min faster cold per variant vs pre-PR (skips gRPC compile).
• Compare a future hot-cache build (same LLAMA_VERSION) against pre-PR — ccache + skipped gRPC stage should bring -gpu-nvidia-cuda-13-llama-cpp from the recent 6h timeout to well under 1h.

Revert plan

Each commit is independently revertable:

• E: revert if l4t-cuda-12 variant has issues — falls back to "those 2 entries aren't mapped" (broken builds).
• D: revert if matrix mapping has issues — backends fall back to empty builder-base-image and Dockerfiles fail because they require it. Means revert C+D together.
• C: same.
• B: revert if the variant Dockerfile refactor has issues — the from-source path comes back. Revert B+C+D+E together.
• A: revert only if builder-base-image input has issues — but this is purely additive, low-risk.

The cleanest safe revert is "revert PR 2 entirely" which restores the from-source variant Dockerfiles. The base images on quay remain harmlessly available for future use.

Assisted-by: Claude:claude-opus-4-7