#include <opennn/network/layers/grouped_query_attention_layer.h>

Inheritance diagram for opennn::GroupedQueryAttention:

Public Member Functions
	GroupedQueryAttention (const Shape &=Shape({0, 0}), Index q_heads=0, Index kv_heads=0, Index head_dim=0, float rope_theta=1000000.0f, float rms_epsilon=1.0e-6f, bool use_qk_norm=true, const string &="grouped_query_attention")

Shape	get_input_shape () const noexcept override

Shape	get_output_shape () const override

Index	get_sequence_length () const

Index	get_hidden () const

Index	get_q_heads () const

Index	get_kv_heads () const

Index	get_head_dim () const

bool	get_use_qk_norm () const

vector< TensorSpec >	get_forward_specs (Index) const override

bool	is_forward_slot_inference_elidable (size_t, Device, Index) const noexcept

bool	uses_compact_inference () const noexcept

ForwardSlotKind	get_forward_slot_kind (size_t slot) const override

void	set (const Shape &, Index, Index, Index, float, float, bool new_use_qk_norm, const string &)

bool	accepts_input_rank (Index rank) const override

bool	uses_sequence_position () const noexcept override

void	apply_input_shape (const Shape &) override

void	on_compute_dtype_changed () override

void	read_JSON_body (const Json *) override

void	write_JSON_body (JsonWriter &) const override

virtual bool	is_forward_slot_inference_elidable (size_t, Device) const noexcept

Public Member Functions inherited from opennn::Layer
virtual	~Layer ()=default

const string &	get_label () const noexcept

const string &	get_name () const

LayerType	get_type () const noexcept

void	set_input_shape (const Shape &new_input_shape)

virtual void	set_output_shape (const Shape &)

void	set_label (string new_label)

Index	get_parameters_number () const

const vector< Operator * > &	get_operators () const noexcept

virtual vector< TensorSpec >	get_parameter_specs () const

virtual vector< TensorSpec >	get_state_specs () const

virtual vector< TensorSpec >	get_backward_specs (Index batch_size) const

virtual size_t	get_recomputable_forward_slot () const noexcept

virtual void	recompute_forward_slot (ForwardPropagation &, size_t)

virtual bool	backward_uses_forward_output () const noexcept

virtual bool	backward_uses_input (size_t) const noexcept

virtual bool	preserves_output_delta_during_backward () const noexcept

virtual bool	allows_input_delta_alias () const noexcept

virtual bool	folds_input_delta_addend (size_t) const noexcept

virtual ActivationFunction	get_output_activation () const

virtual Index	get_sources_number () const noexcept

virtual bool	allows_successors () const noexcept

virtual bool	is_recurrent () const noexcept

virtual bool	skip_for_pre_scaled_input () const noexcept

virtual bool	allows_bf16_input_cast (size_t) const noexcept

Index	get_inputs_number () const noexcept

Index	get_outputs_number () const

virtual void	forward_propagate (ForwardPropagation &forward_propagation, size_t layer, ForwardPropagationMode pass)

virtual void	back_propagate (ForwardPropagation &forward_propagation, BackPropagation &back_propagation, size_t i) const

virtual void	from_JSON (const JsonDocument &)

virtual void	on_loaded ()

virtual void	load_state_from_JSON (const JsonDocument &)

virtual void	to_JSON (JsonWriter &) const

virtual string	write_expression (const vector< string > &, const vector< string > &) const

bool	get_is_trainable () const noexcept

void	set_is_trainable (bool trainable)

Type	get_compute_dtype () const noexcept

Device	get_compute_device () const noexcept

void	set_compute_dtype (Type new_compute_dtype)

void	set_compute_device (Device new_compute_device)

virtual float *	link_states (float *, Device)

float *	link_gradients (float *, Device)

vector< TensorView > &	get_parameter_views ()

const vector< TensorView > &	get_parameter_views () const noexcept

vector< TensorView > &	get_parameter_scales ()

const vector< TensorView > &	get_parameter_scales () const noexcept

vector< Operator::SlotQuantization >	get_parameter_quantization () const

virtual TiedWeight	get_tied_weight () const

virtual void	set_tied_weight (const TiedWeight &tied_weight)

void	redistribute_parameters_to_operators ()

Friends
class	Qwen3

Additional Inherited Members
Protected Types inherited from opennn::Layer
enum	Forward { Input , Output }

Protected Member Functions inherited from opennn::Layer
	Layer (LayerType type, Trainability trainability=Trainability::Trainable)

float *	link_views_to_operators (vector< TensorView > &, float , vector< TensorSpec >(Operator::specs_fn)() const, void(Operator::*link_fn)(span< const TensorView >), Device)

Static Protected Member Functions inherited from opennn::Layer
static bool	refresh_feature_storage (Buffer &storage, bool &dirty, Device device, Index features, Index columns, const function< void(float *)> &fill)

Protected Attributes inherited from opennn::Layer
string	label = "my_layer"

bool	is_trainable = true

Shape	input_shape

Type	compute_dtype = Type::FP32

Type	weights_dtype = Type::FP32

Device	compute_device = Device::CPU

vector< TensorView >	parameters

vector< TensorView >	parameter_scales

vector< TensorView >	states

vector< Operator * >	operators

Constructor & Destructor Documentation

◆ GroupedQueryAttention()

opennn::GroupedQueryAttention::GroupedQueryAttention	(	const Shape &	= Shape({0, 0}),
		Index	q_heads = 0,
		Index	kv_heads = 0,
		Index	head_dim = 0,
		float	rope_theta = 1000000.0f,
		float	rms_epsilon = 1.0e-6f,
		bool	use_qk_norm = true,
		const string &	= "grouped_query_attention" )

Member Function Documentation

◆ accepts_input_rank()

bool opennn::GroupedQueryAttention::accepts_input_rank ( Index rank ) const

inlineoverridevirtual

Reimplemented from opennn::Layer.

◆ apply_input_shape()

void opennn::GroupedQueryAttention::apply_input_shape ( const Shape & )

overridevirtual

Reimplemented from opennn::Layer.

◆ get_forward_slot_kind()

ForwardSlotKind opennn::GroupedQueryAttention::get_forward_slot_kind ( size_t slot ) const

inlineoverridevirtual

Reimplemented from opennn::Layer.

◆ get_forward_specs()

vector< TensorSpec > opennn::GroupedQueryAttention::get_forward_specs ( Index ) const

overridevirtual

Reimplemented from opennn::Layer.

◆ get_head_dim()

Index opennn::GroupedQueryAttention::get_head_dim ( ) const

inline

◆ get_hidden()

Index opennn::GroupedQueryAttention::get_hidden ( ) const

inline

◆ get_input_shape()

Shape opennn::GroupedQueryAttention::get_input_shape ( ) const

inlineoverridevirtualnoexcept

Reimplemented from opennn::Layer.

◆ get_kv_heads()

Index opennn::GroupedQueryAttention::get_kv_heads ( ) const

inline

◆ get_output_shape()

Shape opennn::GroupedQueryAttention::get_output_shape ( ) const

inlineoverridevirtual

Implements opennn::Layer.

◆ get_q_heads()

Index opennn::GroupedQueryAttention::get_q_heads ( ) const

inline

◆ get_sequence_length()

Index opennn::GroupedQueryAttention::get_sequence_length ( ) const

inline

◆ get_use_qk_norm()

bool opennn::GroupedQueryAttention::get_use_qk_norm ( ) const

inline

◆ is_forward_slot_inference_elidable() [1/2]

virtual bool opennn::Layer::is_forward_slot_inference_elidable	(	size_t	,
		Device	) const

inlinevirtualnoexcept

Reimplemented from opennn::Layer.

◆ is_forward_slot_inference_elidable() [2/2]

bool opennn::GroupedQueryAttention::is_forward_slot_inference_elidable	(	size_t	,
		Device	,
		Index	) const

noexcept

◆ on_compute_dtype_changed()

void opennn::GroupedQueryAttention::on_compute_dtype_changed ( )

inlineoverridevirtual

Reimplemented from opennn::Layer.

◆ read_JSON_body()

void opennn::GroupedQueryAttention::read_JSON_body ( const Json * )

overridevirtual

Reimplemented from opennn::Layer.

◆ set()

void opennn::GroupedQueryAttention::set	(	const Shape &	,
		Index	,
		Index	,
		Index	,
		float	,
		float	,
		bool	new_use_qk_norm,
		const string &	)

◆ uses_compact_inference()

bool opennn::GroupedQueryAttention::uses_compact_inference ( ) const

noexcept

◆ uses_sequence_position()

bool opennn::GroupedQueryAttention::uses_sequence_position ( ) const

inlineoverridevirtualnoexcept

Reimplemented from opennn::Layer.

◆ write_JSON_body()

void opennn::GroupedQueryAttention::write_JSON_body ( JsonWriter & ) const

overridevirtual

Reimplemented from opennn::Layer.

Friends And Related Symbol Documentation

◆ Qwen3

friend class Qwen3

friend

Public Member Functions

Friends

Additional Inherited Members

Constructor & Destructor Documentation

◆ GroupedQueryAttention()

Member Function Documentation

◆ accepts_input_rank()

◆ apply_input_shape()

◆ get_forward_slot_kind()

◆ get_forward_specs()

◆ get_head_dim()

◆ get_hidden()

◆ get_input_shape()

◆ get_kv_heads()

◆ get_output_shape()

◆ get_q_heads()

◆ get_sequence_length()

◆ get_use_qk_norm()

◆ is_forward_slot_inference_elidable() [1/2]

◆ is_forward_slot_inference_elidable() [2/2]

◆ on_compute_dtype_changed()

◆ read_JSON_body()

◆ set()

◆ uses_compact_inference()

◆ uses_sequence_position()

◆ write_JSON_body()

Friends And Related Symbol Documentation

◆ Qwen3