mirror of
https://github.com/Helldez/BigMoeOnEdge.git
synced 2026-10-03 03:25:42 +00:00
1d37057fac
Some checks failed
A session opened with --decide answers which of a list of choices the model would pick, read from the next-token distribution after one prefill, with no decode. The state after a shared prefix is kept and restored when the next prefix extends it. Android app: a Choose from options switch. With --prefill-device, a decision is prefilled by the chat turn's placement rule and keeps no prefix state: llama.cpp saves a sequence through KV views that do not follow the moved model state (gate G18g). Also fixes the engine version, stuck at 0.23.0 since 0.24.0. App 0.27.0 (42).
79 lines
4.1 KiB
CMake
79 lines
4.1 KiB
CMake
# BigMoeOnEdge — top-level build.
|
|
#
|
|
# The engine (`bmoe_core`) depends only on its own public headers under include/bmoe
|
|
# and on llama.cpp's PUBLIC C API (llama.h / ggml.h). The serial streaming path never
|
|
# touches llama.cpp internals: MoE expert streaming is driven entirely through the public
|
|
# eval-callback and public gguf accessors, so it tracks upstream with nothing more than a
|
|
# submodule pointer bump — no in-tree patch. The one exception is the optional `--overlap`
|
|
# feature, which needs a per-expert wait point the public API does not expose: the submodule
|
|
# pins a fork branch (`bmoe/expert-ready-hook` on Helldez/llama.cpp) adding one ~25-line
|
|
# readiness hook, detected here as BMOE_HAVE_EXPERT_READY_HOOK. It is zero-cost when absent,
|
|
# the serial path still builds against stock upstream, and it sunsets the moment upstream
|
|
# ships an equivalent callback. See docs/seam.md § 3.
|
|
#
|
|
# llama.cpp is an optional dependency at configure time: when the submodule is absent
|
|
# only the pure-policy code (config validation) compiles, so the scaffold and a subset
|
|
# of CI stay green before the native dependency is fetched.
|
|
cmake_minimum_required(VERSION 3.21)
|
|
|
|
# The version is declared here and nowhere else in the build: the engine reports it (bmoe-cli
|
|
# --version, and the run-parameter preamble of every metrics CSV), so a committed benchmark file
|
|
# says which engine produced it. Keep it in step with CHANGELOG.md and the app's versionName.
|
|
project(bigmoeonedge VERSION 0.27.0 LANGUAGES CXX)
|
|
|
|
set(CMAKE_CXX_STANDARD 17)
|
|
set(CMAKE_CXX_STANDARD_REQUIRED ON)
|
|
set(CMAKE_CXX_EXTENSIONS OFF)
|
|
|
|
if(NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
|
|
set(CMAKE_BUILD_TYPE Release CACHE STRING "" FORCE)
|
|
endif()
|
|
|
|
option(BMOE_BUILD_TESTS "Build byte-identity gates and unit tests" ON)
|
|
option(BMOE_BUILD_CLI "Build the bmoe-cli host tool" ON)
|
|
# Off by default: diagnostics, not part of the product. They need no llama.cpp, so they can be
|
|
# configured and cross-compiled on their own without building the engine.
|
|
option(BMOE_BUILD_TOOLS "Build standalone diagnostic tools (bmoe-iobench)" OFF)
|
|
|
|
# --- Optional native dependency: llama.cpp ----------------------------------------
|
|
# The streaming seam uses only the public C API (llama.h). We additionally link llama.cpp's
|
|
# `common` layer for one thing: chat-template rendering + reasoning parsing (common_chat_*),
|
|
# so prompt formatting and thinking-extraction are template-driven per model instead of
|
|
# hardcoded. Unlike the streaming seam, `common` is NOT a stable API — it can change between
|
|
# upstream versions, so a submodule bump may need the chat glue in runtime.cpp updated (the
|
|
# gates catch it at build time). That trade-off is deliberate; see docs/seam.md.
|
|
set(BMOE_HAVE_LLAMA OFF)
|
|
if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/third_party/llama.cpp/CMakeLists.txt")
|
|
set(BMOE_HAVE_LLAMA ON)
|
|
# Upstream examples/tests/server stay off; build the common utils we link for chat.
|
|
set(LLAMA_BUILD_TESTS OFF CACHE BOOL "" FORCE)
|
|
set(LLAMA_BUILD_EXAMPLES OFF CACHE BOOL "" FORCE)
|
|
set(LLAMA_BUILD_SERVER OFF CACHE BOOL "" FORCE)
|
|
set(LLAMA_BUILD_COMMON ON CACHE BOOL "" FORCE)
|
|
set(LLAMA_CURL OFF CACHE BOOL "" FORCE) # common's model-download path unused
|
|
# The prefill-device gate (G16) needs a second device to move weights onto, and the one every
|
|
# host has is a loopback rpc-server fronting the CPU. A test fixture, not a feature: host test
|
|
# builds only (releases build without tests), no front-end accepts an RPC endpoint, and it is
|
|
# not forced: -DGGML_RPC=OFF still wins, and G16 then reports itself skipped.
|
|
if(BMOE_BUILD_TESTS AND NOT ANDROID)
|
|
set(GGML_RPC ON CACHE BOOL "ggml: use RPC")
|
|
endif()
|
|
add_subdirectory(third_party/llama.cpp EXCLUDE_FROM_ALL)
|
|
endif()
|
|
|
|
add_subdirectory(core)
|
|
|
|
if(BMOE_BUILD_CLI AND BMOE_HAVE_LLAMA)
|
|
add_subdirectory(cli)
|
|
endif()
|
|
|
|
if(BMOE_BUILD_TESTS AND BMOE_HAVE_LLAMA)
|
|
enable_testing()
|
|
add_subdirectory(tests)
|
|
endif()
|
|
|
|
if(BMOE_BUILD_TOOLS)
|
|
add_subdirectory(tools)
|
|
endif()
|
|
|
|
message(STATUS "BigMoeOnEdge: llama.cpp present = ${BMOE_HAVE_LLAMA}")
|