MAJEEngine and SDK
The core provides tensors, model layers, training, evaluation and serialization. The Python/C++ SDK uses the same native engine.
Current capabilities
- ModelsImplemented
- Tensors, dense networks, CNNs, recurrent layers and graph models.
- Training and inferenceImplemented
- Train supported models, evaluate results and run inference.
- CheckpointsImplemented
- Save models and resume supported training runs.
- CPU and CUDAScoped support
- CPU Tensor/CNN training is the default. CUDA support depends on the model, precision and execution mode.
Stable support is limited to qualified configurations. Other implemented paths remain Experimental.
Python and C++ SDK
The SDK is implemented and admitted for experimental use. Some capabilities still lack full validation.
Ask about the SDKC++
An API for native applications.
Python
Bindings to the C++ engine.
Availability varies by package and backend. The SDK as a whole remains Experimental.
Precision, autograd and support limits
Typed FP32 programs, selected mixed precision and CPU INT8 inference are available within defined domains. Automatic differentiation and ahead-of-time specialization are Experimental and opt-in.
Support depends on the operation, backend, precision, execution mode and workload limits. Native engine qualifications do not automatically apply to the SDK.
MAJE Lab runs reproducible checks and measurements. Published results apply to the workloads and configurations tested.