Mila::Dnn::Model< TDeviceType, TPrecision > Class Template Reference

Inheritance diagram for Mila::Dnn::Model< TDeviceType, TPrecision >:

Public Types
using	NetworkType = Network<TDeviceType, TPrecision>

Public Member Functions
	Model (const Model &)=delete
	Model (Model &&)=default
DeviceId	getDeviceId () const noexcept
	The device this model runs on.
MemoryStats	getMemoryStats () const
	Current memory allocation breakdown for this model.
RuntimeMode	getRuntimeMode () const noexcept
	The runtime mode this model was constructed for.
std::size_t	getScratchHighWaterBytes () const
	Context-owned scratch device memory, in bytes, at its high-water mark.
bool	isEval () const noexcept
	True if this model is currently in eval sub-state.
bool	isInferenceMode () const noexcept
	True if this model was constructed for inference.
bool	isTrainingMode () const noexcept
	True if this model was constructed for training.
Model &	operator= (const Model &)=delete
Model &	operator= (Model &&)=default
void	setEval (bool eval)
	Toggle eval sub-state for this model.
virtual std::string	toString () const =0
	Human-readable summary of this model's configuration.
void	train ()
	Run the training loop for this model.

Protected Member Functions
	Model (std::unique_ptr< NetworkType > network, RuntimeMode runtime_mode)
	Construct with a fully built network and runtime mode.
virtual void	onTraining ()=0
	Training loop hook – derived class owns the implementation.

Protected Attributes
std::unique_ptr< NetworkType >	network_
	The owned Network instance.

Constructor & Destructor Documentation

Model()

template<DeviceType TDeviceType, TensorDataType TPrecision>

Mila::Dnn::Model< TDeviceType, TPrecision >::Model ( std::unique_ptr< NetworkType > network, RuntimeMode runtime_mode )

inlineexplicitprotected

Construct with a fully built network and runtime mode.

Called by derived class constructors only. The network must already be built and have parameters loaded before this constructor is called.

Parameters

network	Fully built and loaded Network.
runtime_mode	Inference or Training – immutable after construction.

Member Function Documentation

getRuntimeMode()

template<DeviceType TDeviceType, TensorDataType TPrecision>

RuntimeMode Mila::Dnn::Model< TDeviceType, TPrecision >::getRuntimeMode ( ) const

inlinenoexcept

The runtime mode this model was constructed for.

Immutable after construction. Governs which public API methods are valid.

getScratchHighWaterBytes()

template<DeviceType TDeviceType, TensorDataType TPrecision>

std::size_t Mila::Dnn::Model< TDeviceType, TPrecision >::getScratchHighWaterBytes ( ) const

inline

Context-owned scratch device memory, in bytes, at its high-water mark.

Deliberately NOT folded into getMemoryStats(): that reports what components allocate, and this buffer belongs to the shared execution context, is allocated lazily during forward passes, and is sized by whichever operation needed the most at any point. Callers wanting the true resident total add the two.

Zero before the first forward pass, and on any backend that allocates no scratch.

isEval()

template<DeviceType TDeviceType, TensorDataType TPrecision>

bool Mila::Dnn::Model< TDeviceType, TPrecision >::isEval ( ) const

inlinenoexcept

True if this model is currently in eval sub-state.

For RuntimeMode::Inference models always returns true – inference models never compute gradients. For RuntimeMode::Training models reflects the last setEval() call.

isInferenceMode()

template<DeviceType TDeviceType, TensorDataType TPrecision>

bool Mila::Dnn::Model< TDeviceType, TPrecision >::isInferenceMode ( ) const

inlinenoexcept

True if this model was constructed for inference.

The model-family inference API (e.g. generate()) is valid. train() will throw.

isTrainingMode()

template<DeviceType TDeviceType, TensorDataType TPrecision>

bool Mila::Dnn::Model< TDeviceType, TPrecision >::isTrainingMode ( ) const

inlinenoexcept

True if this model was constructed for training.

train() is valid. The model-family inference API will throw.

onTraining()

template<DeviceType TDeviceType, TensorDataType TPrecision>

virtual void Mila::Dnn::Model< TDeviceType, TPrecision >::onTraining ( )

protectedpure virtual

Training loop hook – derived class owns the implementation.

Called by train() after precondition enforcement. The derived class has total control over data loading, optimizer construction, loss computation, backward pass, checkpointing, and sampling.

Pure virtual – a model declaring RuntimeMode::Training must provide a training loop.

Implemented in Mila::Dnn::GemmaModel< TDeviceType, TPrecision >, Mila::Dnn::GptModel< TDeviceType, TPrecision >, and Mila::Dnn::LlamaModel< TDeviceType, TPrecision >.

setEval()

template<DeviceType TDeviceType, TensorDataType TPrecision>

void Mila::Dnn::Model< TDeviceType, TPrecision >::setEval ( bool eval )

inline

Toggle eval sub-state for this model.

When eval is true, the forward pass runs without gradients, dropout is disabled, and batch norm uses running statistics. When eval is false, the full training pass is restored.

Cascades through Network to every Component and Operation in the graph via their onEvalChanging() hooks.

Only valid on models constructed with RuntimeMode::Training. Inference-mode models are always in eval state by definition.

Parameters

eval	true to enter eval sub-state, false to restore training.

Exceptions

std::runtime_error

if called on a RuntimeMode::Inference model.

toString()

template<DeviceType TDeviceType, TensorDataType TPrecision>

virtual std::string Mila::Dnn::Model< TDeviceType, TPrecision >::toString ( ) const

pure virtual

Human-readable summary of this model's configuration.

Implemented in Mila::Dnn::GemmaModel< TDeviceType, TPrecision >, Mila::Dnn::GptModel< TDeviceType, TPrecision >, and Mila::Dnn::LlamaModel< TDeviceType, TPrecision >.

train()

template<DeviceType TDeviceType, TensorDataType TPrecision>

void Mila::Dnn::Model< TDeviceType, TPrecision >::train ( )

inline

Run the training loop for this model.

Enforces RuntimeMode::Training precondition then delegates entirely to onTraining(). The derived class owns the loop – data loading, optimizer, loss, checkpointing, and sampling are all derived class concerns.

Exceptions

std::runtime_error

if called on an Inference-mode model.

Member Data Documentation

network_

template<DeviceType TDeviceType, TensorDataType TPrecision>

std::unique_ptr<NetworkType> Mila::Dnn::Model< TDeviceType, TPrecision >::network_

protected

The owned Network instance.

Accessible to derived classes for model-specific operations not covered by the base class API.

---

The documentation for this class was generated from the following file:

• Mila/Src/Dnn/Core/Model.ixx