engine-versions.env

# Pinned sources for the bundled engine, shared by the wheel build
# (tools/wheel-build/build_llama_server.sh) and CI's engine-helper install.
# Bump deliberately; a change here invalidates the cached engine binaries.
#
# ENGINE_LLAMA_CPP_VERSION: the llama-cpp-python release tag whose vendored
# llama.cpp commit llama-server is built from. Re-run the Metal/CPU/GPU
# self-check matrix when bumping. llama-cpp-python is only a BUILD-TIME
# source; lilbee no longer depends on it at runtime.
ENGINE_LLAMA_CPP_VERSION=0.3.30

# The two Go engine helpers bundled alongside llama-server, built from source
# (deterministic, no release-asset-name guessing).
ENGINE_LLAMA_SWAP_VERSION=v223
ENGINE_GGUF_PARSER_REF=v0.25.0

# ENGINE_GO_VERSION: the Go toolchain that builds those helpers, consumed by
# ci.yml, build-multigpu.yml and the QA pod scripts. llama-swap's go.mod sets
# the floor (v223 wants >= 1.26.1); the builds run with GOTOOLCHAIN=local so a
# go.mod bump past this pin fails loudly instead of silently downloading one.
ENGINE_GO_VERSION=1.26.5

# ENGINE_VULKAN_SDK_VERSION: the LunarG SDK the Linux vulkan cells build against.
# It carries glslc and the SPIR-V headers, neither of which apt provides on the
# jammy runner. Shared by install_gpu_toolkit.sh, release.yml and
# flatpak-validate.yml so the three cannot drift.
ENGINE_VULKAN_SDK_VERSION=1.3.296.0

# ENGINE_ROCM_VERSION: installed, compiled against, and bundled by the rocm cell.
# Below 6.1 llama.cpp refuses to build and 6.1 fails on ggml-cuda/mma.cuh. A bump
# can drop GPU targets; cmake_args.sh filters those against the toolchain.
ENGINE_ROCM_VERSION=7.2.0