fix(build): link llama-common-base so CUDA/SYCL source builds link

[SnowCheetos]

Overview

Fixes the CUDA/SYCL link failure seen in v0.17.5 (run 28915721077). It was not OOM/timeout — the source builds now compile fully and fail at link:

rust-lld: error: undefined symbol: llama_build_info()
                 undefined symbol: llama_build_number()
                 undefined symbol: llama_commit() / llama_compiler() / llama_build_target()

(Good news: the resilience change from #44 worked — v0.17.5 still published the macOS, Linux CPU, and Linux Vulkan binaries despite these two failing.)

Root cause

Since b9837, llama.cpp compiles build-info into a dedicated static archive llama-common-base that libllama-common depends on but that cmake does not install. So build.rs's install-dir scan links libllama-common.a (which references llama_build_number() etc.) but never libllama-common-base.a (which defines them) → undefined symbols. Prebuilt targets are unaffected: their libllama-common.so already contains build-info.

Fix

• build.rs: link libllama-common-base.a explicitly, emitted after the install libs so it follows llama-common on the link line (static-link order: the referencer must precede the definer).
• bridge: drop the #ifndef LLAMA_CPP_PREBUILT fallback globals (LLAMA_BUILD_NUMBER = 0, …). Those were a stand-in for when build-info wasn't linked; now that we link the real archive, keeping them would cause duplicate-symbol errors. build-info comes from libllama-common in both modes.

Verification (against a real b9837 checkout)

• nm libllama-common-base.a → defines llama_build_number (T), llama_build_info (T), LLAMA_BUILD_NUMBER (D), …
• nm libllama-common.a → references them (U).
• Bridge compiles in both source and prebuilt modes after removing the fallbacks.

Caveat: I can't run the CUDA/SYCL toolchains (or even a full CPU source build — this box lacks openblas-devel) locally, so this resolves the confirmed undefined symbols but can't rule out a further downstream link issue on those two backends. #44's resilience change backstops that — the other three targets still ship regardless.

Worth considering

This is the 4th sequential source-build break since the b9837 bump. Vulkan already covers NVIDIA + AMD + Intel GPUs and passes CI. It may be worth dropping the dedicated CUDA/SYCL release targets rather than continuing to chase these — happy to send that as a one-line matrix change instead.

[Copilot]

Pull request overview

Resolves the CUDA/SYCL from-source build link failure caused by missing llama.cpp build-info symbols by ensuring the new llama-common-base static archive is linked, and by removing now-conflicting fallback symbol definitions from the C++ bridge.

Changes:

• Update build.rs to explicitly link llama-common-base (emitted after install-dir libs to preserve static link order).
• Remove non-prebuilt fallback LLAMA_BUILD_* global definitions from the C++ bridge to avoid duplicate-symbol collisions once build-info is properly linked.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
crates/llama-sys/build.rs	Explicitly links llama-common-base from the build tree to satisfy build-info symbols missing from the install-dir scan.
crates/llama-sys/src/autocommit_common_bridge.cpp	Removes fallback build-info globals that would collide with real definitions once llama-common-base is linked.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.