Model import paths

Artifact	Path	Runtime	Allocation contract	Status
Mamba / supported flow checkpoint	convert/convert.py → .safetensors	FlowEdge Core	Zero after setup	Supported
LeRobot diffusion_pusht directory	policy_inspect.py → converter → flowedge-lerobot	Core + companion	Zero in native inference after setup	Supported
Fixed-shape one-I/O ONNX policy	flowedge-onnx	Optional ONNX Runtime	External runtime; measure separately	Supported adapter
ExecuTorch .pte	Future companion adapter	ExecuTorch	External runtime; measure separately	Planned

Use the smallest path that preserves your policy. Native conversion is the only path covered by FlowEdge’s zero-hot-path-allocation contract. The ONNX adapter validates one fixed-shape float32 input/output pair at setup, has no dependency from Core or Relay, and reports operator/provider failures instead of silently falling back.

python convert/policy_inspect.py policy_dir --json
python -m pip install -e 'integrations/onnx[runtime]'

The CI fixture exports a deterministic PyTorch policy and checks its ONNX Runtime output through OnnxAdapter.run_into. Repeat that parity check for your model before a control loop. Dynamic shapes, multiple I/O, and unsupported operators need a model-specific adapter or native converter rather than a hidden graph runtime.