GPT-2 style transformer (decoder-only) for autoregressive token prediction. More...

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

Public Types
using	ComponentPtr = typename NetworkBase::ComponentPtr
using	EncoderType = Lpe<TDeviceType, dtype_t::INT32, TPrecision>
using	LayerNormType = LayerNorm<TDeviceType, TPrecision>
using	LinearType = Linear<TDeviceType, TPrecision>
using	MR = typename DeviceTypeTraits<TDeviceType>::memory_resource
using	NetworkBase = LanguageNetwork<TDeviceType, TPrecision>
using	TensorType = Tensor<TPrecision, MR>
using	TokenIndexType = Tensor<dtype_t::INT32, MR>
using	TransformerBlockType = GptBlock<TDeviceType, TPrecision>
Public Types inherited from Mila::Dnn::LanguageNetwork< TDeviceType, TPrecision >
using	MR = typename DeviceTypeTraits<TDeviceType>::memory_resource
using	NetworkBase = Network<TDeviceType, TPrecision>
using	StageProbe = std::function<void( std::string_view stage, const TensorType& value )>
	Observer called with each intermediate activation during prefill.
using	TensorType = Tensor<TPrecision, MR>
using	TokenIndexType = Tensor<TensorDataType::INT32, MR>
Public Types inherited from Mila::Dnn::Network< TDeviceType, TPrecision >
using	ComponentPtr = typename CompositeBase::ComponentPtr
using	CompositeBase = CompositeComponent<TDeviceType, TPrecision>
using	MR = typename DeviceTypeTraits<TDeviceType>::memory_resource
Public Types inherited from Mila::Dnn::CompositeComponent< TDeviceType, TPrecision >
using	ComponentBase = Component<TDeviceType, TPrecision>
using	ComponentPtr = std::shared_ptr<Component<TDeviceType, TPrecision>>

Public Member Functions
	GptTransformer (const std::string &name, const GptConfig &config, DeviceId device_id)
	Construct Gpt type transformer.
TokenIndexType &	backward (const TokenIndexType &input, const TensorType &output_grad) override
TensorType &	decode (const TokenIndexType &input, dim_t position) override
	Inference-only single-token decode pass.
TensorType &	forward (const TokenIndexType &input) override
IExecutionContext *	getExecutionContext () const
MemoryStats	getMemoryStats () const override
	Return the current memory allocation breakdown for this component.
ModelType	getModelType () const
void	loadParameters (PretrainedModelReader &reader, bool strict)
	Initialize this transformer's components from a GPT-2 checkpoint.
TensorType &	prefill (const TokenIndexType &input) override
	Inference prefill – process full prompt and return last-token logits.
std::string	toString () const override
	Generate a human-readable description.
void	zeroGradients () override
	Clear all model-owned gradients for this component.
Public Member Functions inherited from Mila::Dnn::LanguageNetwork< TDeviceType, TPrecision >
	LanguageNetwork (const std::string &name)
virtual TokenIndexType &	backward (const TokenIndexType &input, const TensorType &output_grad)=0
	Full backward pass (training).
virtual TensorType &	decode (const TokenIndexType &input, dim_t position)=0
	Inference decode – single-token autoregressive step.
virtual TensorType &	forward (const TokenIndexType &input)=0
	Full-sequence forward pass.
virtual TensorType &	prefill (const TokenIndexType &input)=0
	Inference prefill – process full prompt and populate the KV cache.
virtual TensorType &	prefillFrom (const TokenIndexType &input, dim_t start_offset)
	Chunked prefill starting at an absolute position (prompt-prefix reuse).
virtual bool	rewindKvCache (dim_t position)
	Rewind the KV caches to position for prompt-prefix reuse (PromptCaching.md).
virtual void	setStageProbe (StageProbe probe)
	Install a stage probe, or clear it by passing an empty function.
Public Member Functions inherited from Mila::Dnn::Network< TDeviceType, TPrecision >
	Network (const std::string &name)
	Construct network (context managed by derived class).
template<typename TOptimizer, typename TConfig>
std::shared_ptr< TOptimizer >	createOptimizer (const TConfig &config)
	Create and configure an optimizer for this network's parameters.
DeviceId	getDeviceId () const noexcept override
	Get the compute device for this composite.
IExecutionContext *	getExecutionContext () const
	Public access to the network's shared execution context.
const ComponentType	getType () const override
	Get the component type identifier.
void	load (ModelArchive &archive, SerializationMode mode)
	Restore this network's parameters from an archive.
void	save (ModelArchive &archive, SerializationMode mode) const
	Save network to archive.
void	synchronize () override
	Synchronize all child components.
std::string	toString () const override
	Generate a human-readable description.
Public Member Functions inherited from Mila::Dnn::CompositeComponent< TDeviceType, TPrecision >
	CompositeComponent (CompositeComponent &&) noexcept=default
	CompositeComponent (const CompositeComponent &)=delete
	CompositeComponent (const std::string &name)
	Construct composite component with name.
CompositeComponent &	addComponent (ComponentPtr component)
	Add a pre-constructed child component (chainable).
size_t	childCount () const noexcept
	Get the number of direct children.
void	clearComponents ()
	Clear all child components.
ComponentPtr	findComponent (const std::string &path) const
	Resolve a dot-separated component path within this composite.
ComponentPtr	getComponent (const std::string &name) const
	Retrieve a direct child component by name.
const std::vector< ComponentPtr > &	getComponents () const
	Get all child components in insertion order.
std::vector< ITensor * >	getGradients () const override
	Get all parameter gradients from all children.
std::vector< ITensor * >	getParameters () const override
	Get all parameters from all children.
bool	hasChildren () const noexcept
	Check if this composite has any children.
bool	hasComponent (const std::string &name) const
	Check if a named child component exists.
CompositeComponent &	operator= (CompositeComponent &&) noexcept=default
CompositeComponent &	operator= (const CompositeComponent &)=delete
dim_t	parameterCount () const override
	Count parameters across all children.
bool	removeComponent (const std::string &name)
void	saveFlatTensors (Serialization::SafeTensorsWriter &writer, const std::string &prefix, Serialization::TensorSavePass pass) const override
	Recurse into children, extending the flat dotted prefix.
ComponentPtr	tryFindComponent (const std::string &path) const
	Try to resolve a dot-separated component path within this composite.
Public Member Functions inherited from Mila::Dnn::Component< TDeviceType, TPrecision >
	Component (const std::string &name)
	Construct component with required name identifier.
virtual void	build (const BuildContext &context) final
	Build the component with the provided BuildContext (canonical overload).
const std::string	getName () const
	Get the component's name identifier.
virtual std::vector< std::string >	getParameterNames () const
	List all available parameter names for this component.
virtual MemoryStats	getRequiredMemory (const BuildContext &context) const
	Report what build( context ) would allocate, without allocating it.
TrainingMode	getTrainingMode () const noexcept
	The current runtime behavioral mode of this Component.
virtual bool	isBuilt () const final
	Returns true if build() has completed successfully.
virtual void	loadParameter (const std::string &, const Serialization::ITensorBlob &)
	Load a parameter from serialized tensor data.
void	setTrainingMode (TrainingMode mode)
	Set the runtime behavioral mode for this Component.

Static Public Member Functions
static dim_t	buildSequenceLengthFromArchive (const ModelArchive &archive)
	Sequence length the archive was built with, or 0 if it recorded none.
static GptConfig	configFromArchive (const ModelArchive &archive)
	Reconstruct the GptConfig that produced an archive.
static std::unique_ptr< GptTransformer< TDeviceType, TPrecision > >	fromPretrained (const std::filesystem::path &model_path, std::size_t batch_size, std::size_t seq_length, DeviceId device_id=DeviceId{ TDeviceType, 0 }, bool strict=true)
Static Public Member Functions inherited from Mila::Dnn::Component< TDeviceType, TPrecision >
static constexpr DeviceType	getDeviceType ()
	Compile-time device type for this component instance.
static constexpr TensorDataType	getPrecision () noexcept
	Compile-time tensor precision for this component instance.

Protected Member Functions
void	load_ (ModelArchive &archive, SerializationMode) override
	Validate that the archive describes the network that was built.
void	onBuilding (const BuildContext &context) override
	Hook invoked by build() to allocate component buffers.
void	onTrainingModeChanging (TrainingMode training_mode) override
	Hook invoked when training mode is about to change.
void	save_ (ModelArchive &archive, SerializationMode) const override
	Hook for concrete classes to save type-specific state.
Protected Member Functions inherited from Mila::Dnn::Network< TDeviceType, TPrecision >
void	verifyArchitectureCompatibility (const PretrainedMetadata &metadata)
	Verify that imported model is compatible with network architecture.
Protected Member Functions inherited from Mila::Dnn::CompositeComponent< TDeviceType, TPrecision >
template<typename TComponent>
std::shared_ptr< TComponent >	getComponentAs (const std::string &name) const
	Retrieve a typed child component by name.
void	onExecutionContextSet () override
	Hook invoked after ExecutionContext is set.
virtual void	optimize ()
	Virtual hook for graph optimization after construction.
void	requireSerializableParameters () const override
	No-op override: a composite names no parameters of its own.
Protected Member Functions inherited from Mila::Dnn::Component< TDeviceType, TPrecision >
IExecutionContext *	getExecutionContext () const
	Get the shared execution context.
bool	hasExecutionContext () const noexcept
	Check if execution context has been set.
template<TensorDataType TParameterPrecision, typename TMemoryResource>
void	loadParameterFromBlob (const std::string &param_name, const Serialization::ITensorBlob &blob, Tensor< TParameterPrecision, TMemoryResource > &target, const shape_t &expected_shape)
	Load a tensor blob into a parameter tensor with validation.
template<TensorDataType TParameterPrecision, typename TMemoryResource>
void	saveParameterToArchive (ModelArchive &archive, const std::string &parameter_name, const Tensor< TParameterPrecision, TMemoryResource > &parameter) const
	Write one parameter tensor into the archive under "tensors/<name>".
template<TensorDataType TParameterPrecision, typename TMemoryResource>
void	saveParameterToWriter (Serialization::SafeTensorsWriter &writer, const std::string &flat_name, const Tensor< TParameterPrecision, TMemoryResource > &parameter, Serialization::TensorSavePass pass) const
	Drive one parameter through one pass of a flat safetensors save.
void	setExecutionContext (IExecutionContext *context)
	Set the execution context for this component.

Additional Inherited Members
Protected Types inherited from Mila::Dnn::Component< TDeviceType, TPrecision >
using	HostStagingMemoryResource
	Host memory a device-resident parameter stages through.
Protected Attributes inherited from Mila::Dnn::Component< TDeviceType, TPrecision >
BuildContext	build_context_ { shape_t{ 1 }, RuntimeMode::Training }
	The BuildContext stored at build time.

Detailed Description

template<DeviceType TDeviceType, TensorDataType TPrecision>
requires PrecisionSupportedOnDevice<TPrecision, TDeviceType>
class Mila::Dnn::GptTransformer< TDeviceType, TPrecision >

GPT-2 style transformer (decoder-only) for autoregressive token prediction.

Template parameters:

TDeviceType: device type (Cpu/Cuda)
TPrecision: tensor precision

Constructor & Destructor Documentation

◆ GptTransformer()

template<DeviceType TDeviceType, TensorDataType TPrecision>

Mila::Dnn::GptTransformer< TDeviceType, TPrecision >::GptTransformer	(	const std::string &	name,
		const GptConfig &	config,
		DeviceId	device_id )

inlineexplicit

Construct Gpt type transformer.

Parameters

name	Network name
config	GPT transformer configuration
device_id	Device identifier for execution

Exceptions

std::invalid_argument on invalid config or device mismatch

Member Function Documentation

◆ buildSequenceLengthFromArchive()

template<DeviceType TDeviceType, TensorDataType TPrecision>

dim_t Mila::Dnn::GptTransformer< TDeviceType, TPrecision >::buildSequenceLengthFromArchive ( const ModelArchive & archive )

inlinestatic

Sequence length the archive was built with, or 0 if it recorded none.

save_() writes the build geometry only when the network was built, so a caller restoring a checkpoint uses this to rebuild at the same geometry rather than guessing.

◆ configFromArchive()

template<DeviceType TDeviceType, TensorDataType TPrecision>

GptConfig Mila::Dnn::GptTransformer< TDeviceType, TPrecision >::configFromArchive ( const ModelArchive & archive )

inlinestatic

Reconstruct the GptConfig that produced an archive.

Reads the config block save_() wrote and hands it to GptConfig::fromMetadata, which owns the key names and their fallbacks. The caller then constructs and builds the network before calling Network::load() to restore weights – the config must exist before the network can, which is why this is separate from the load traversal rather than a step inside it.

Note the scope: save_() runs at the archive root (Network::save calls it after saveComponentGraph has popped its scopes), so "transformer_meta.json" is a root-level entry.

Parameters

archive Archive opened for reading.

Returns: Config equivalent to the one the archive was saved from.

Exceptions

std::runtime_error if the archive carries no transformer metadata.

◆ decode()

template<DeviceType TDeviceType, TensorDataType TPrecision>

TensorType & Mila::Dnn::GptTransformer< TDeviceType, TPrecision >::decode	(	const TokenIndexType &	input,
		dim_t	position )

inlineoverride

Inference-only single-token decode pass.

Mirrors forward() exactly except each transformer block is driven via decode() rather than forward(). Each block's decode() delegates to attn_->decode() for the attention step – Attention decides internally whether to use the fast KV cache path or fall back to forward(). All other components in each block use forward() unchanged.

The encoder (token + position embeddings) and final LayerNorm + LM head are identical to forward() – only the block traversal differs.

Precondition: forward() must have been called at least once (prefill) before decode() is called. Attention internally manages cache state – no explicit initializeKVCache / resetKVCache needed here.

Calling forward() again after decode() steps automatically resets the KV cache and begins a new prefill session.

Parameters

input	Single-token input [B, 1] token indices.
position	Current sequence position (0-based).

Returns: Reference to logits tensor [B, 1, vocab_size].

◆ getMemoryStats()

template<DeviceType TDeviceType, TensorDataType TPrecision>

MemoryStats Mila::Dnn::GptTransformer< TDeviceType, TPrecision >::getMemoryStats ( ) const

inlineoverridevirtual

Return the current memory allocation breakdown for this component.

Reflects allocations at the moment of the call. The returned stats naturally track the component lifecycle:

After construction – nothing; construction allocates none After build( Inference ) – parameters + T=1 state buffers After build( Training ) – parameters + T=full state buffers After setTrainingMode( Train ) – parameters + state + gradients

For CompositeComponent and Network, the returned stats are the recursive aggregate of all child components.

May be called at any time – no lifecycle preconditions.

Returns: MemoryStats reflecting current allocations.

Implements Mila::Dnn::Component< TDeviceType, TPrecision >.

◆ load_()

template<DeviceType TDeviceType, TensorDataType TPrecision>

void Mila::Dnn::GptTransformer< TDeviceType, TPrecision >::load_	(	ModelArchive &	archive,
		SerializationMode	)

inlineoverrideprotectedvirtual

Validate that the archive describes the network that was built.

The config was consumed by the factory before construction, so there is nothing to restore here – only to check. A mismatch means the caller built from a different config than the one that produced the archive, which would otherwise surface as a shape error deep inside a component's loadParameter.

Reimplemented from Mila::Dnn::Network< TDeviceType, TPrecision >.

◆ loadParameters()

template<DeviceType TDeviceType, TensorDataType TPrecision>

void Mila::Dnn::GptTransformer< TDeviceType, TPrecision >::loadParameters	(	PretrainedModelReader &	reader,
		bool	strict )

inline

Initialize this transformer's components from a GPT-2 checkpoint.

Delegates to small helpers that load checkpoint blobs and apply them to the encoder, per-layer blocks, and final layer-norm.

Load parameters (weights and biases) from an already-opened PretrainedModelReader

Separated from fromPretrained to allow flexibility in weight loading

◆ onBuilding()

template<DeviceType TDeviceType, TensorDataType TPrecision>

void Mila::Dnn::GptTransformer< TDeviceType, TPrecision >::onBuilding ( const BuildContext & context )

inlineoverrideprotectedvirtual

Hook invoked by build() to allocate component buffers.

Receives the stored BuildContext. Implementations must use config.allocationSeqLen() when sizing output buffers – this is the single call that makes Inference and Training allocate the correct buffer sizes automatically without per-component logic.

// Example -- Linear component:
shape_t out_shape =
{
    config.batchSize(),
    config.allocationSeqLen(),   // 1 for Inference, T for Training
    config_.getOutputFeatures()
};
output_ = std::make_unique<TensorType>( device, out_shape,
    this->getName() + ".output" );

The default implementation forwards to the legacy onBuilding( const shape_t& ) overload for backwards compatibility. New components should override this overload directly.

Note: Do not call build() or onBuilding() from within this hook.; Implementations should either succeed fully or leave no partial state, as a failed build() may be retried.

Takes the build-time configuration; use its allocationSeqLen() to obtain the correct output buffer sequence dimension.

Reimplemented from Mila::Dnn::Component< TDeviceType, TPrecision >.

◆ onTrainingModeChanging()

template<DeviceType TDeviceType, TensorDataType TPrecision>

void Mila::Dnn::GptTransformer< TDeviceType, TPrecision >::onTrainingModeChanging ( TrainingMode training_mode )

inlineoverrideprotectedvirtual

Hook invoked when training mode is about to change.

Propagates the new mode to all child components. The hook runs with the Component's training mutex held; it MUST NOT call setTrainingMode().

Parameters

training_mode New training mode (Normal or Eval)

Reimplemented from Mila::Dnn::CompositeComponent< TDeviceType, TPrecision >.

◆ prefill()

template<DeviceType TDeviceType, TensorDataType TPrecision>

TensorType & Mila::Dnn::GptTransformer< TDeviceType, TPrecision >::prefill ( const TokenIndexType & input )

inlineoverride

Inference prefill – process full prompt and return last-token logits.

Populates the KV cache across all transformer blocks by running the full prompt through encoder + blocks via forward(). Then extracts only the last token's representation for the final LayerNorm + LM head, avoiding the T=1 output buffer overflow that forward() would cause on those components.

Unlike LlamaTransformer::prefill(), GPT does not need chunked prefill or explicit position offsets (no RoPE). The full sequence is processed in a single pass.

Parameters

input Full prompt token indices [B, T].

Returns: Logits for the last token [B, 1, vocab_size].

◆ save_()

template<DeviceType TDeviceType, TensorDataType TPrecision>

void Mila::Dnn::GptTransformer< TDeviceType, TPrecision >::save_	(	ModelArchive &	archive,
		SerializationMode	mode ) const

inlineoverrideprotectedvirtual

Hook for concrete classes to save type-specific state.

REQUIRED override for concrete networks. Must write:

Type identifier (e.g., "type": "MnistClassifier")
Configuration parameters (batch_size, architecture constants)
Shape metadata (for validation during Load())

This metadata enables the concrete class's Load() method to reconstruct the network.

Example implementation:

void save_(ModelArchive& archive, SerializationMode mode) const override
{
    json meta;
    meta["type"] = "MnistClassifier";  // Type identifier for runtime dispatch
    meta["batch_size"] = batch_size_;
    meta["input_shape"] = leading_shape_;
    // ... other configuration
    archive.writeJson("network/classifier_meta.json", meta);
}

Parameters

archive	Archive to write to
mode	Serialization mode (passed from save())

Implements Mila::Dnn::Network< TDeviceType, TPrecision >.

◆ toString()

template<DeviceType TDeviceType, TensorDataType TPrecision>

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

inlineoverridevirtual

Generate a human-readable description.

Returns: String representation showing children

Reimplemented from Mila::Dnn::CompositeComponent< TDeviceType, TPrecision >.

◆ zeroGradients()

template<DeviceType TDeviceType, TensorDataType TPrecision>

void Mila::Dnn::GptTransformer< TDeviceType, TPrecision >::zeroGradients ( )

inlineoverridevirtual

Clear all model-owned gradients for this component.

Default implementation is a no-op. Composite components should override to recurse to children. Leaf components should override to zero their parameter and activation gradients using device-aware helpers.

Reimplemented from Mila::Dnn::Component< TDeviceType, TPrecision >.

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

Mila/Src/Dnn/Components/Transformers/Gpt/GptTransformer.ixx

Public Types

Public Member Functions

Static Public Member Functions

Protected Member Functions

Additional Inherited Members

Detailed Description

Constructor & Destructor Documentation

◆ GptTransformer()

Member Function Documentation

◆ buildSequenceLengthFromArchive()

◆ configFromArchive()

◆ decode()

◆ getMemoryStats()

◆ load_()

◆ loadParameters()

◆ onBuilding()

◆ onTrainingModeChanging()

◆ prefill()

◆ save_()

◆ toString()

◆ zeroGradients()