Skip to content
cortexkitPublic

About

Local AI inference for CortexKit: embeddings, rerank, micro-LLM one-shots across CUDA/Metal/ANE/Vulkan/CPU — certification-gated engines behind one subc surface.

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

1,502 Commits

Folders and files

Repository files navigation

Synapse

Synapse is a local inference service on the subc bus. It provides embedding, reranking and small-model generation through the synapse module, with explicit model installation, stable numerical identity, bounded scheduling and durable result pages. Owned engines and supervised workers isolate inference from the management surface. Configured remote profiles require explicit identity opt-in. A missing local backend produces a typed refusal, not an automatic CPU fallback.

Workspace crates

Production and supporting workspace members live under crates/. A worker's presence in source does not imply that a release ships it; the catalog and release matrix determine supported backends and published assets.

Member under crates/ Responsibility
synapse-build-stamp Shared clean/dirty Git build provenance.
synapse-core Protocol, errors, fingerprints, tokenizer, cache and scheduler types.
synapse-native-probe Safe native platform-profile probes.
synapse-certify Schema-3 certification records, self-checks and offline validation.
synapse-certify-runner Non-shipped hardware evidence producer with isolated run state.
synapse-engine-owned Owned Metal embedding/reranking and decode kernels.
synapse-engine-owned/owned-decode-worker Owned-decode supervisor, protocol and crash budget.
synapse-engine-cuda Owned CUDA engine and hardware-floor checks.
synapse-module Management operations, model cache, durable jobs and worker host; ck-synapse.
synapse-opctl Operator client; ck-synapse-opctl.
synapse-worker-llama Supervised GGUF worker.
synapse-worker-ane Rust launcher and Swift Core ML worker.
synapse-worker-ane-direct Direct Neural Engine compilation, admission and inference worker.
synapse-worker-decode Supervised owned-generation worker.
synapse-worker-cuda Owned CUDA worker.
synapse-worker-vulkan Owned Vulkan embedding/reranking worker.

Build and test

Use the committed Cargo.lock. The workspace requires Rust 1.89 or newer and native tools appropriate to the selected package. Apple Metal builds need full Xcode, not only Command Line Tools; verify xcrun --find metal on macOS.

From the repository root:

cargo build --locked -p synapse-module -p synapse-opctl
cargo build --locked --no-default-features -p synapse-worker-cuda -p synapse-worker-vulkan
cargo check --locked -p synapse-module --features test-support
cargo test --locked -p synapse-module --features test-support --lib --test it --test skeleton_e2e --test wire_contract
cargo test --locked -p synapse-certify --lib --test records
cargo fmt --all --check
bash scripts/public-banlist-check.sh
bash scripts/check-cortexkit-layout.sh

The feature-off GPU workers above are test fixtures, not GPU certification. Hardware and checkpoint-dependent tests are separate from portable tests; a skip is not a hardware pass. See development verification and backend certification for those boundaries.

Documentation

ARCHITECTURE.md and STRUCTURE.md remain root code/session maps. Public certification artifacts are tracked because release tooling reads them; internal research and operational records are not part of the public documentation set.

About

Local AI inference for CortexKit: embeddings, rerank, micro-LLM one-shots across CUDA/Metal/ANE/Vulkan/CPU — certification-gated engines behind one subc surface.

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages