A pruned fork of Strata (engine 0.1.44.1): the C++/CUDA/HIP inference engine for Qwen3.8-Flash-Next and Qwen3.6-35B-A3B mixture-of-experts models, kept buildable and close to upstream so it can follow upstream later.
The upstream installer, Python server, web app, Intel SYCL and Apple Metal ports, Docker packaging, benchmark-result archives and user documentation are removed. See FORK_NOTES.md for exactly what was removed, what is kept, and what is deferred.
Requires CMake 3.24+, a C++20 compiler, and a CUDA toolkit (or AMD HIP, or CPU-only). First check out
the vendored llama.cpp submodule (pinned at tag b11624, commit af0579c) — the build fetches
nothing on its own:
git submodule update --init
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DSTRATA_ENABLE_CUDA=ON -DSTRATA_BUILD_TESTS=ON
cmake --build build -j
ctest --test-dir build
ggml, the MMQ kernel sources and ggml-common.h all come from the pinned submodule
(STRATA_GGML_DIR, -DSTRATA_GGML_DIR=... overrides it).
src/,include/— the engine (core, kernels, prefill, speculative decoding, plan).tests/— engine tests (built whenSTRATA_BUILD_TESTS=ON).tools/— model preparation (pack, draft head, expert profiles) and small parity helpers.data/— small runtime data (draft vocabulary, expert profiles).ref/— reference GGUF loader used to validate the engine's reading of GGUF.bench/micro/— micro-benchmarks and parity harnesses referenced by the build.
There is no serving/API layer yet; the engine is driven through its CLI binaries
(strata, strata-plan, strata-q36, ...). A llama.cpp-based frontend is planned as a
separate phase (see FORK_NOTES.md).