Meshflow

Cerebras-inspired spatial inference engine.

Meshflow compiles small neural networks into spatial execution plans and runs them on a simulated 2D mesh of processing elements (PEs). Weights are resident in PE-local memory, activations flow along explicit routes.

Prerequisites

• Python 3.12+
• Rust (stable toolchain)
• uv for Python dependency management

Setup

uv sync
uv run maturin develop

Verify installation

uv run python -c "import meshflow; print(meshflow.__version__)"

Running tests

# All tests
./scripts/test.sh

# Python tests only
uv run pytest tests/python

# Rust tests only
cargo test -p mesh_runtime

Lint and format

./scripts/lint.sh

Development server

uv run uvicorn meshflow.api.server:app --reload

Rebuild Rust extension

After changing Rust code:

uv run maturin develop

Project structure

meshflow/
├── python/meshflow/      # Python package (compiler, API, tools)
├── crates/mesh_runtime/  # Rust runtime (mesh simulator core)
├── tests/python/         # Python tests
├── tests/rust/           # Rust integration tests
├── artifacts/            # Generated compiled programs and traces
└── scripts/              # Dev workflow scripts