fix(build): disable LLAMA_BUILD_APP/UI so the CUDA/SYCL source build links

[SnowCheetos]

Overview

Follow-up to #42. That PR fixed 3 of the 5 release targets and bumped the llama.cpp submodule to b9837, which resolved the bridge API drift on the two from-source targets (CUDA, SYCL). But those two then failed at a new point — v0.17.3 Release run 28913925408:

third_party/llama.cpp/app/llama.cpp:1:10: fatal error: build-info.h: No such file or directory
[3%] Building CXX object app/CMakeFiles/llama-app.dir/llama.cpp.o

Root cause

b9837 introduced two new CMake options that both default ON for a standalone (top-level) build:

option(LLAMA_BUILD_APP "llama: build the unified binary" ${LLAMA_STANDALONE})   # new
option(LLAMA_BUILD_UI  "llama: build the embedded Web UI for server" ON)        # new
...
if (LLAMA_BUILD_APP)
    add_subdirectory(app)
endif()

build.rs disables TESTS/EXAMPLES/TOOLS/SERVER but not these newer targets, so cmake built the app/llama-app binary — which #includes a generated build-info.h it can't find. We only need the libraries (libllama, libggml*, libcommon) for FFI, so the app binary should never be built.

Prebuilt targets (CPU/Vulkan/macOS) are unaffected — they download .sos and never run the cmake source build. This only bit the two from-source backends.

Fix

Add -DLLAMA_BUILD_APP=OFF -DLLAMA_BUILD_UI=OFF to the cmake configure in build_from_source. Older llama.cpp releases that lack these options ignore the flags harmlessly.

Verification

Reproduced and validated locally against a b9837 checkout with the exact build.rs flags:

Configure	app subdir added?
current main flags (no LLAMA_BUILD_APP)	yes — reproduces the failure
+ -DLLAMA_BUILD_APP=OFF	no — target excluded ✓

• All flags recognized (no "unused variable" warnings).
• llama, ggml, ggml-base, common targets still present.
• libllama.a builds cleanly past the point where CI failed.

CI note: the Release matrix only runs on a v* tag, so this won't be exercised by this PR's checks — it needs a tag (or a workflow_dispatch trigger) to confirm green end-to-end.

[Copilot]

Pull request overview

This PR adjusts the llama-sys from-source CMake configuration to prevent llama.cpp’s newly-added application/UI targets from being built when compiling CUDA/SYCL backends from source, ensuring the build only produces the libraries needed for FFI.

Changes:

• Disable llama.cpp’s LLAMA_BUILD_APP target during from-source builds to avoid compiling app/ (and its generated-header dependency).
• Disable LLAMA_BUILD_UI during from-source builds to avoid the embedded UI target behavior.
• Add in-file rationale documenting why these options are explicitly turned off.

---

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