SonicBoom

SonicBoom is a native tensor runtime — a generic interpreter (v0) that executes neural-network graphs through native C++ backends, with a first-class Guile binding.

Architecture

Four layers with strictly downward dependencies:

Layer 3    C ABI / Guile / other bindings
Layer 2    SonicBoom stable runtime interface (nt::)
Layer 1    native-torch implementation + adapter
Layer 0    CPU / CUDA

libsonicboom.so is the published core library (built SHARED); native-torch and MLIR are statically linked into it (position-independent code). The stable binary-compatibility boundary is the C ABI (capi/); the C++ Layer 2 interface is source-level only.

Guile AI framework

The (sonicboom …) modules provide a minimal PyTorch/Keras-like model API:

1. Define / compose — (sonicboom nn): linear, relu, sequential.
2. Train — (sonicboom model): make-model + model-train! drive native-torch reverse-mode autograd (linear/relu/mse/SGD) through the C ABI; parameters live as native tensors, not Scheme copies.
3. Export — model-export! writes a self-contained model.sx + weights.bin artifact via sb_export_model.
4. Independent inference — standalone C++ sonicboom::Model::load runs the S-Expr → MLIR path with no Guile or Python.

Training (native-torch autograd) and deployment (S-Expr → MLIR) are two engines that agree on structure, parameter naming, shapes, and math semantics. See design/guile-ai-framework-progress.md for the step-by-step record.

Build

C++23 with g++-13. MLIR/LLVM are pinned and installed to build/llvm-mlir-install via tools/mlir/build-mlir.sh.

cmake -S . -B build
cmake --build build

Test

ctest --test-dir build --output-on-failure

runs the full suite — 40 tests (36 C++ suites + 4 Guile suites) — in one pass.