Gemma 4 transformer (decoder-only) for autoregressive inference. More...

Inheritance diagram for Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >:

Public Types
using	ComponentPtr = typename NetworkBase::ComponentPtr
using	DecoderLayerType = IDecoderLayer<TDeviceType, TPrecision>
using	GlobalBlockType = GemmaBlock<TDeviceType, TPrecision, true, TWeightQuantization, NoKvCompression>
using	LmHeadLinearType = Linear<TDeviceType, TPrecision, TableQuantizationPolicy>
using	LocalBlockType = GemmaBlock<TDeviceType, TPrecision, false, TWeightQuantization, TKvCachePolicy>
using	MR = typename DeviceTypeTraits<TDeviceType>::memory_resource
using	NetworkBase = LanguageNetwork<TDeviceType, TPrecision>
using	RmsNormType = RmsNorm<TDeviceType, TPrecision>
using	TableQuantizationPolicy
using	TensorType = Tensor<TPrecision, MR>
using	TokenEmbeddingType = TokenEmbedding<TDeviceType, dtype_t::INT32, TPrecision, TableQuantizationPolicy>
using	TokenIndexType = Tensor<dtype_t::INT32, MR>
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
	GemmaTransformer (const std::string &name, const GemmaConfig &config, DeviceId device_id)
TokenIndexType &	backward (const TokenIndexType &, const TensorType &) override
TensorType &	decode (const TokenIndexType &input, dim_t position) override
TensorType &	forward (const TokenIndexType &) override
IExecutionContext *	getExecutionContext () const
MemoryStats	getMemoryStats () const override
	Return the current memory allocation breakdown for this component.
ModelType	getModelType () const
MemoryStats	getRequiredMemory (const BuildContext &context) const override
	What build( context ) would allocate for the whole model, without allocating.
void	loadParameters (PretrainedModelReader &reader)
dim_t	parameterCount () const override
	Count parameters across all children.
TensorType &	prefill (const TokenIndexType &input) override
TensorType &	prefillFrom (const TokenIndexType &input, dim_t start_offset) override
	Chunked prefill starting at an absolute position (prompt-prefix reuse).
void	resetKVCache ()
bool	rewindKvCache (dim_t position) override
	Rewind every layer's KV cache to position for prefix reuse.
void	setStageProbe (typename NetworkBase::StageProbe probe) override
	Install a stage probe on the prefill path.
std::string	toString () const override
	Generate a human-readable description.
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).
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
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.
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.
virtual void	zeroGradients ()
	Clear all model-owned gradients for this component.

Protected Member Functions
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 >
virtual void	load_ (ModelArchive &archive, SerializationMode mode) override
	Hook for concrete classes to validate type-specific state on load.
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
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 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, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>
requires PrecisionSupportedOnDevice<TPrecision, TDeviceType>
class Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >

Gemma 4 transformer (decoder-only) for autoregressive inference.

Graph: TokenEmbedding -> GemmaBlock x N (heterogeneous local/global) -> RmsNorm -> Linear (lm_head). The embedding sqrt(d) scale and the final logit softcap are handled by the converter and the sampler respectively (see the file header).

Member Typedef Documentation

◆ TableQuantizationPolicy

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

using Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >::TableQuantizationPolicy

Initial value:

std::conditional_t<

TWeightQuantization::kIsQuantized, PerChannelFp8<>, NoWeightQuant>

Mila::Dnn::Quant::Weight::NoWeightQuant

Definition Policies.ixx:24

Mila::Dnn::Quant::Weight::PerChannelFp8

Definition Policies.ixx:48

Member Function Documentation

◆ getMemoryStats()

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

MemoryStats Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >::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 >.

◆ getRequiredMemory()

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

MemoryStats Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >::getRequiredMemory ( const BuildContext & context ) const

inlineoverridevirtual

What build( context ) would allocate for the whole model, without allocating.

Mirrors onBuilding(): resolve the prefill chunk first, then recurse with the same per-child contexts, then add the transformer's own pooled buffers and apply the two no-double-count corrections. See Specifications/MemoryFootprint.md.

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

◆ onBuilding()

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

void Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >::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, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

void Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >::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 >.

◆ parameterCount()

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

dim_t Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >::parameterCount ( ) const

inlineoverridevirtual

Count parameters across all children.

Returns: Total number of trainable parameters

Exceptions

std::runtime_error if called before build()

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

◆ prefillFrom()

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

TensorType & Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >::prefillFrom	(	const TokenIndexType &	input,
		dim_t	start_offset )

inlineoverride

Chunked prefill starting at an absolute position (prompt-prefix reuse).

input is the FULL prompt tensor, so the token index and the absolute position coincide; chunking simply starts at start_offset instead of 0. Positions [0, start_offset) must already be resident in the KV caches (rewindKvCache). start_offset must lie inside the prompt so at least one position is prefilled and the returned last-position logits are fresh.

◆ rewindKvCache()

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

bool Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >::rewindKvCache ( dim_t position )

inlineoverridevirtual

Rewind every layer's KV cache to position for prefix reuse.

All-or-nothing from the caller's perspective: returns true only when every layer accepted. On false the caller falls back to a full prefill, which positionally overwrites all caches – so a partial rewind (some layers moved, a bounded ring refused) needs no cleanup.

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

◆ save_()

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

void Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >::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 >.

◆ setStageProbe()

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

void Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >::setStageProbe ( typename NetworkBase::StageProbe probe )

inlineoverridevirtual

Install a stage probe on the prefill path.

Empty function clears it.

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

◆ toString()

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

std::string Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >::toString ( ) const

inlineoverridevirtual

Generate a human-readable description.

Returns: String representation showing children

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

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

Mila/Src/Dnn/Components/Transformers/Gemma/Gemma.ixx

Public Types

Public Member Functions

Protected Member Functions

Additional Inherited Members

Detailed Description

Member Typedef Documentation

◆ TableQuantizationPolicy

Member Function Documentation

◆ getMemoryStats()

◆ getRequiredMemory()

◆ onBuilding()

◆ onTrainingModeChanging()

◆ parameterCount()

◆ prefillFrom()

◆ rewindKvCache()

◆ save_()

◆ setStageProbe()

◆ toString()