d7a78599e6
Benchmarked BUILD_SHARED_LIBS ON vs OFF: - Build time identical (~55s both on 192-core B200) - Static: 75MB self-contained binary, no LD_LIBRARY_PATH needed - Shared: 7.4MB + .so deps, needs lib path management - Both work correctly Optimizations applied to LlamaCppBuilder: - BUILD_SHARED_LIBS=OFF: static binary, simpler deployment - ccache: 27x faster rebuilds (55s -> 2s with warm cache) - Ninja generator: parallel builds - CMAKE_CUDA_ARCHITECTURES: build only for detected GPU arch - CMAKE_CUDA_FLAGS=--threads=0: multi-threaded CUDA compilation - GGML_NATIVE=ON: native CPU optimizations - LLAMA_BUILD_TESTS=OFF, LLAMA_BUILD_EXAMPLES=OFF: skip unused targets - LLAMA_BUILD_SERVER=ON: build only what we need |
||
|---|---|---|
| .. | ||
| assets | ||
| auth | ||
| core | ||
| loggers | ||
| models | ||
| plugins | ||
| requirements | ||
| routes | ||
| state | ||
| tests | ||
| utils | ||
| __init__.py | ||
| colab.py | ||
| main.py | ||
| run.py | ||