Feature Request: C/C++ code generation backend for embedded deployment

[akhodeir]

Hi Hummingbird team,

First of all, thank you for the excellent work on converting tree-based models into tensor / GEMM representations.

I’m currently using Hummingbird to convert XGBoost and decision tree models into GEMM-based implementations for fast inference.

This works very well for CPU/GPU inference, but for embedded systems with limited resources, the current deployment options are difficult to use:
• Torch / TorchScript runtime is too heavy
• saving .pt files is not practical
• ONNX Runtime is also too large for MCU-class targets
• we need deterministic static memory usage
• firmware environments often require pure C/C++ source code

Proposed feature

It would be extremely valuable to add a backend such as:

convert(model, backend="c")
convert(model, backend="cpp")

that generates:
• static const weight matrices
• threshold vectors
• branch masks
• leaf projection matrices
• optional fixed-point quantized arrays
• lightweight GEMM kernels
• pure C/C++ inference functions
• no external runtime dependency

This would allow deployment on:
• ARM Cortex-M
• DSPs
• microcontrollers
• bare-metal firmware
• RTOS systems
• safety-critical embedded environments