Model.ixx File Reference

Abstract base class for all Mila models. More...

#include <memory>
#include <string>
#include <stdexcept>
#include <format>
import Compute.DeviceId;
import Dnn.TensorDataTypeTraits;
import Compute.DeviceType;
import Dnn.TensorDataType;
import Dnn.RuntimeMode;
import Dnn.TensorTypes;
import Dnn.Component;
import Dnn.Network;

Classes
class	Mila::Dnn::Model< TDeviceType, TPrecision >

Namespaces
namespace	Mila
	Mila main API namespace.
namespace	Mila::Dnn

Detailed Description

Abstract base class for all Mila models.

Model defines the universal contract across all model families — language models, image classifiers, regression models, and any future model type.

Architecture

Model sits at the top of the Mila DNN pipeline:

Component — leaf node, shape-driven buffer allocation CompositeComponent — structural aggregation, cascades BuildConfig Network — graph topology, forward/backward Model — RuntimeMode, lifecycle, universal API boundary LanguageModel — generate(), sampling, EOS, vocabulary GptModel — GPT-specific factory and config LlamaModel — LLaMA-specific factory and config ImageClassifier— classify(), top-k predictions

RuntimeMode

A Model is constructed for either Inference or Training — immutable after construction. The mode governs which public API methods are valid:

Mode	Valid
Inference	model-family inference API
Training	train() → onTraining() hook

Training

train() enforces the RuntimeMode::Training precondition then delegates entirely to the pure virtual onTraining() hook. The derived class owns the training loop — data loading, optimizer, loss, backward pass, checkpointing, and sampling are all derived class concerns.

Threading

Not thread-safe. External synchronization required if shared.