Gemma 4 compatible inference model. More...

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

Public Types
using	GemmaSlidingKvPolicy = Quant::KvCache::SlidingWindowKvCache
	KV policy for Gemma's LOCAL (sliding) layers.
using	ModelBase = LanguageModel<TDeviceType, TPrecision>
using	MR = typename DeviceTypeTraits<TDeviceType>::memory_resource
using	TensorType = Tensor<TPrecision, MR>
using	TokenIndexType = Tensor<dtype_t::INT32, MR>
Public Types inherited from Mila::Dnn::LanguageModel< TDeviceType, TPrecision >
using	Base = Model<TDeviceType, TPrecision>
Public Types inherited from Mila::Dnn::Model< TDeviceType, TPrecision >
using	NetworkType = Network<TDeviceType, TPrecision>

Public Member Functions
	GemmaModel (const GemmaModel &)=delete
	GemmaModel (GemmaModel &&)=default
dim_t	contextLength () const noexcept
	Deployment context length: the KV-cache depth the network was built with.
const GemmaModelConfig &	getModelConfig () const noexcept
	Deployment configuration (context length, weight-quant, kv-compression) this model was loaded with.
const GemmaConfig &	getNetworkConfig () const noexcept
	Architecture/network configuration (read from the checkpoint metadata).
GemmaModel &	operator= (const GemmaModel &)=delete
GemmaModel &	operator= (GemmaModel &&)=default
std::string	toString () const override
	Human-readable summary of this model's configuration.
Public Member Functions inherited from Mila::Dnn::LanguageModel< TDeviceType, TPrecision >
	LanguageModel (const LanguageModel &)=delete
	LanguageModel (LanguageModel &&)=default
std::vector< std::string >	componentPaths () const
	Every observable component path, for choosing a pattern or diagnosing a zero.
GenerateStatus	generate (std::span< const int32_t > prompt_tokens, const std::function< void(int32_t)> &on_token, const GenerateParams &params={}, std::stop_token stop={})
	Generate tokens from a prompt, streaming each through on_token.
size_t	observe (std::string_view pattern, Compute::ComputePassMask passes, Compute::ActivationObserver sink)
	Watch activations flowing through matching components on matching passes.
LanguageModel &	operator= (const LanguageModel &)=delete
LanguageModel &	operator= (LanguageModel &&)=default
void	save (const std::filesystem::path &path) const
	Write this model's live weights as a safetensors weights file.
void	seedSampler (uint64_t seed)
	Seed the sampler's RNG for reproducible generation.
void	stopObserving ()
	Detach every observer and clear the sink. A probe that outlives its question costs.
std::string	weightQuantizationScheme () const
	The scheme name save records for these weights, e.g.
Public Member Functions inherited from Mila::Dnn::Model< TDeviceType, TPrecision >
	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.
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.
void	train ()
	Run the training loop for this model.

Static Public Member Functions
static GemmaConfig	configFromMetadata (const WeightsMetadata &metadata)
	The network geometry a load of this file builds.
static DeploymentFootprint	getDeploymentFootprint (const std::filesystem::path &path, const GemmaModelConfig &model_config, DeviceId device_id=DeviceId{ TDeviceType, 0 })
	The same prediction, plus how this deployment would chunk its prefill.
static MemoryStats	getRequiredMemory (const std::filesystem::path &path, const GemmaModelConfig &model_config, DeviceId device_id=DeviceId{ TDeviceType, 0 })
	What loading this checkpoint at this context length would cost in VRAM.
static std::unique_ptr< GemmaModel< TDeviceType, TPrecision > >	load (const std::filesystem::path &path, const GemmaModelConfig &model_config, DeviceId device_id=DeviceId{ TDeviceType, 0 })
	Load Gemma 4 from a Mila weights file.

Protected Member Functions
dim_t	maxSequenceLength () const noexcept override
GenerateStatus	onGenerating (std::span< const int32_t > prompt_tokens, const std::function< void(int32_t)> &on_token, const GenerateParams &params, std::stop_token stop) override
	Prefill + decode implementation hook.
void	onTraining () override
	Training loop hook – derived class owns the implementation.
dim_t	vocabSize () const noexcept override
Protected Member Functions inherited from Mila::Dnn::LanguageModel< TDeviceType, TPrecision >
	LanguageModel (std::unique_ptr< LanguageModelNetwork< TDeviceType, TPrecision > > network, RuntimeMode runtime_mode, Serialization::WeightsMetadata source_metadata={}, WeightQuantization weight_quantization=WeightQuantization::None, int fp4_group_size=128)
int32_t	awaitSampledToken ()
	Block until the last enqueueSampleNext()'s token id is host-visible.
void	enqueueSampleNext (const TensorType &logits, TokenTensor &token_out, const SamplingParams &params)
	Enqueue a sampling step without waiting for the host readback.
const LanguageModelNetwork< TDeviceType, TPrecision > &	getNetwork () const noexcept
LanguageModelNetwork< TDeviceType, TPrecision > &	getNetwork () noexcept
int32_t	sampleNext (const TensorType &logits, TokenTensor &token_out, const SamplingParams &params)
	Sample the next token from a logits row on the device.
Protected Member Functions inherited from Mila::Dnn::Model< TDeviceType, TPrecision >
	Model (std::unique_ptr< NetworkType > network, RuntimeMode runtime_mode)
	Construct with a fully built network and runtime mode.

Additional Inherited Members
Protected Types inherited from Mila::Dnn::LanguageModel< TDeviceType, TPrecision >
using	MR = typename DeviceTypeTraits<TDeviceType>::memory_resource
using	TensorType = Tensor<TPrecision, MR>
using	TokenTensor = Tensor<TensorDataType::INT32, MR>
Protected Attributes inherited from Mila::Dnn::LanguageModel< TDeviceType, TPrecision >
int	fp4_group_size_ { 128 }
	The FP4 group the live weights were built at; meaningful only when they are FP4.
Serialization::WeightsMetadata	source_metadata_
	The loaded weights file's metadata, carried so save can write it back verbatim.
WeightQuantization	weight_quantization_ { WeightQuantization::None }
	What the live weights are, not what the source file was.
Protected Attributes inherited from Mila::Dnn::Model< TDeviceType, TPrecision >
std::unique_ptr< NetworkType >	network_
	The owned Network instance.

Detailed Description

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

Gemma 4 compatible inference model.

Owns a loaded, built GemmaTransformer and drives the prefill + KV-cache decode two-phase generation loop. Construction is only possible via load(); the network is always built, weights-loaded, and in inference mode when generation runs.

Thread safety: not thread-safe; external synchronization required if shared.

Member Typedef Documentation

◆ GemmaSlidingKvPolicy

template<DeviceType TDeviceType, TensorDataType TPrecision>

using Mila::Dnn::GemmaModel< TDeviceType, TPrecision >::GemmaSlidingKvPolicy = Quant::KvCache::SlidingWindowKvCache

KV policy for Gemma's LOCAL (sliding) layers.

Bounded sliding-window ring (SlidingWindowKvCache.md Phase 3): their cache is sized to the window working set instead of the full context. Strictly a memory optimization – tokens are identical to the full cache. GLOBAL (full-attention) layers are always NoKvCompression, hardwired in GemmaTransformer. Flip this alias to NoKvCompression to A/B the footprint against the full-context sliding cache.

Class scope rather than per-function so the load and footprint paths cannot be pointed at different policies – that would make a model report a figure for a cache it does not build.

Member Function Documentation

◆ configFromMetadata()

template<DeviceType TDeviceType, TensorDataType TPrecision>

GemmaConfig Mila::Dnn::GemmaModel< TDeviceType, TPrecision >::configFromMetadata ( const WeightsMetadata & metadata )

inlinestatic

The network geometry a load of this file builds.

Public so a caller that constructs blocks itself – the layer-streamed parity harness – uses the geometry a real load would, rather than a second reading of the metadata.

◆ getDeploymentFootprint()

template<DeviceType TDeviceType, TensorDataType TPrecision>

DeploymentFootprint Mila::Dnn::GemmaModel< TDeviceType, TPrecision >::getDeploymentFootprint	(	const std::filesystem::path &	path,
		const GemmaModelConfig &	model_config,
		DeviceId	device_id = DeviceId{ TDeviceType, 0 } )

inlinestatic

The same prediction, plus how this deployment would chunk its prefill.

One graph construction answers both, because they are two readings of the same arithmetic – the chunk is resolved on the way to sizing the activation workspaces getRequiredMemory reports, and was previously discarded there.

The second half is what a caller choosing a context length needs and memory alone cannot tell it: the largest context that fits can be one where the chunk has walked down to its floor, because a longer context leaves less of the free memory for the chunk. See Specifications/MemoryFootprint.md section 11.

Exceptions

std::invalid_argument	on device type mismatch or zero context length.
std::runtime_error	on an unreadable artifact or unsupported quantization.

◆ getRequiredMemory()

template<DeviceType TDeviceType, TensorDataType TPrecision>

MemoryStats Mila::Dnn::GemmaModel< TDeviceType, TPrecision >::getRequiredMemory	(	const std::filesystem::path &	path,
		const GemmaModelConfig &	model_config,
		DeviceId	device_id = DeviceId{ TDeviceType, 0 } )

inlinestatic

What loading this checkpoint at this context length would cost in VRAM.

Reads the artifact header for geometry and constructs the graph, then reports what build() would allocate – without building it, without reading a weight, and therefore without needing the device to have room. Answers before a multi-gigabyte download and for hardware the caller does not own.

Returns measurements only. Whether a given headroom is too tight is a deployment policy and belongs to the adaptor, not here; on Windows in particular WDDM oversubscribes rather than failing, so "fits" is not a property the runtime can decide. See Specifications/MemoryFootprint.md.

Exceptions

std::invalid_argument	on device type mismatch or zero context length.
std::runtime_error	on an unreadable artifact or unsupported quantization.

◆ load()

template<DeviceType TDeviceType, TensorDataType TPrecision>

std::unique_ptr< GemmaModel< TDeviceType, TPrecision > > Mila::Dnn::GemmaModel< TDeviceType, TPrecision >::load	(	const std::filesystem::path &	path,
		const GemmaModelConfig &	model_config,
		DeviceId	device_id = DeviceId{ TDeviceType, 0 } )

inlinestatic

Load Gemma 4 from a Mila weights file.

The model_config carries the deployment decisions (context length, weight quantization, KV-cache compression); every architectural parameter is read from the checkpoint metadata.

Parameters

path	Path to the Gemma weights file.
model_config	Deployment configuration for this load.
device_id	Target device; must match TDeviceType.

Returns: Inference-ready GemmaModel.

Exceptions

std::invalid_argument	on device type mismatch or zero context length.
std::runtime_error	on load failure or unsupported quantization.

◆ maxSequenceLength()

template<DeviceType TDeviceType, TensorDataType TPrecision>

dim_t Mila::Dnn::GemmaModel< TDeviceType, TPrecision >::maxSequenceLength ( ) const

inlineoverrideprotectedvirtualnoexcept

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

◆ onGenerating()

template<DeviceType TDeviceType, TensorDataType TPrecision>

GenerateStatus Mila::Dnn::GemmaModel< TDeviceType, TPrecision >::onGenerating	(	std::span< const int32_t >	prompt_tokens,
		const std::function< void(int32_t)> &	on_token,
		const GenerateParams &	params,
		std::stop_token	stop )

inlineoverrideprotectedvirtual

Prefill + decode implementation hook.

Derived classes own the full autoregressive generation loop. on_token must be called for every generated token except EOS. stop.stop_requested() must be checked on each decode step and generation must abort early when signalled, returning the GenerateStatus that reflects why the loop stopped.

Parameters

prompt_tokens	Input token ids.
on_token	Per-token callback.
params	Per-call generation parameters (loop bound + sampling).
stop	Stop token for cooperative cancellation.

Returns: Why generation stopped.

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

◆ onTraining()

template<DeviceType TDeviceType, TensorDataType TPrecision>

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

inlineoverrideprotectedvirtual

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.

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

◆ toString()

template<DeviceType TDeviceType, TensorDataType TPrecision>

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

inlineoverridevirtual

Human-readable summary of this model's configuration.

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

◆ vocabSize()

template<DeviceType TDeviceType, TensorDataType TPrecision>

dim_t Mila::Dnn::GemmaModel< TDeviceType, TPrecision >::vocabSize ( ) const

inlineoverrideprotectedvirtualnoexcept

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

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

Mila/Src/Dnn/Models/Gemma/GemmaModel.ixx

Public Types

Public Member Functions

Static Public Member Functions

Protected Member Functions

Additional Inherited Members

Detailed Description

Member Typedef Documentation

◆ GemmaSlidingKvPolicy

Member Function Documentation

◆ configFromMetadata()

◆ getDeploymentFootprint()

◆ getRequiredMemory()

◆ load()

◆ maxSequenceLength()

◆ onGenerating()

◆ onTraining()

◆ toString()

◆ vocabSize()