CUDA implementation of the Encoder operation for transformer models. More...

Inheritance diagram for Mila::Dnn::Compute::CudaEncoderOp< TInput, TOutput >:

[legend]

Collaboration diagram for Mila::Dnn::Compute::CudaEncoderOp< TInput, TOutput >:

[legend]

Public Types
using	MR = typename CudaDevice::MR

using	OperationBase = UnaryOperation< DeviceType::Cuda, int, TOutput >

Public Types inherited from Mila::Dnn::Compute::UnaryOperation< TDeviceType, TInput, TOutput >
using	MR = std::conditional_t< TDeviceType==DeviceType::Cuda, CudaMemoryResource, HostMemoryResource >
	Memory resource type based on device type.

Public Member Functions
	CudaEncoderOp (const EncoderConfig &config)
	Constructs a new CUDA Encoder operation with the default device context.

	CudaEncoderOp (std::shared_ptr< DeviceContext > context, const EncoderConfig &config)
	Constructs a new CUDA Encoder operation with a specific device context.

void	backward (const Tensor< int, MR > &input, const Tensor< TOutput, MR > &output, const Tensor< TOutput, MR > &output_gradient, const std::vector< std::shared_ptr< Tensor< TOutput, MR > > > &parameters, std::vector< std::shared_ptr< Tensor< TOutput, MR > > > &parameter_gradients, Tensor< int, MR > &input_gradient, const OperationAttributes &properties, const std::vector< std::shared_ptr< Tensor< TOutput, MR > > > &output_state) const
	Performs the backward pass of the Encoder operation.

void	forward (const Tensor< int, MR > &input, const std::vector< std::shared_ptr< Tensor< TOutput, MR > > > &parameters, const OperationAttributes &properties, Tensor< TOutput, MR > &output, std::vector< std::shared_ptr< Tensor< TOutput, MR > > > &output_state) const override
	Performs the forward pass of the Encoder operation on CUDA.

std::string	getName () const override
	Gets the name of this operation.

Public Member Functions inherited from Mila::Dnn::Compute::UnaryOperation< TDeviceType, TInput, TOutput >
	UnaryOperation (OperationType operation_type)
	Constructs a UnaryOperation with the specified operation type.

	UnaryOperation (OperationType operation_type, std::shared_ptr< DeviceContext > context)
	Constructs a UnaryOperation with the specified operation type and device context.

virtual	~UnaryOperation ()=default
	Virtual destructor for proper cleanup of derived classes.

virtual void	backward (const Tensor< TInput, MR > &grad, const std::vector< std::shared_ptr< Tensor< TOutput, MR > > > &parameters, std::vector< std::shared_ptr< Tensor< TOutput, MR > > > &output_grads) const
	Executes the backward pass of a unary operation.

virtual void	backward (const Tensor< TInput, MR > &input, const Tensor< TOutput, MR > &output_grad, const std::vector< std::shared_ptr< Tensor< TOutput, MR > > > &parameters, std::vector< std::shared_ptr< Tensor< TOutput, MR > > > &parameter_grads, Tensor< TInput, MR > &input_grad, const OperationAttributes &properties, const std::vector< std::shared_ptr< Tensor< TOutput, MR > > > &output_state) const
	Executes the comprehensive backward pass of a unary operation.

virtual void	forward (const Tensor< TInput, MR > &input, const std::vector< std::shared_ptr< Tensor< TOutput, MR > > > &parameters, const OperationAttributes &properties, Tensor< TOutput, MR > &output, std::vector< std::shared_ptr< Tensor< TOutput, MR > > > &output_state) const =0
	Executes the forward pass of a unary operation.

Public Member Functions inherited from Mila::Dnn::Compute::OperationBase< TDeviceType, TInput1, TInput2, TOutput >
	OperationBase (OperationType operation_type, std::shared_ptr< DeviceContext > context)
	Constructs an OperationBase object with a specific device context and compute precision.

virtual	~OperationBase ()=default
	Virtual destructor for the OperationBase class.

std::shared_ptr< DeviceContext >	getDeviceContext () const
	Gets the device context associated with this operation.

DeviceType	getDeviceType () const
	Gets the device type for this operation.

OperationType	getOperationType () const
	Gets the operation type enumeration value.

Private Attributes
EncoderConfig	config_
	Configuration for the encoder operation.

Detailed Description

template<typename TInput, typename TOutput = TInput>
requires ValidFloatTensorTypes<TInput,TOutput>
class Mila::Dnn::Compute::CudaEncoderOp< TInput, TOutput >

CUDA implementation of the Encoder operation for transformer models.

This class provides a CUDA-based implementation of the Encoder operation, which performs token embedding lookups and positional embedding additions. It transforms discrete token IDs into continuous vector representations by combining:

Token embeddings from a learned vocabulary table (wte)
Positional embeddings that encode sequence position information (wpe)

The implementation is optimized for NVIDIA GPUs using CUDA for high-performance computation, supporting both integer and half-precision floating-point operations.

Template Parameters

int	The data type of the input tensor elements (typically uint16_t or int for token IDs).
TDataType	The data type used for computation and output (typically half or float).

Member Typedef Documentation

◆ MR

template<typename TInput , typename TOutput = TInput>

using Mila::Dnn::Compute::CudaEncoderOp< TInput, TOutput >::MR = typename CudaDevice::MR

◆ OperationBase

template<typename TInput , typename TOutput = TInput>

using Mila::Dnn::Compute::CudaEncoderOp< TInput, TOutput >::OperationBase = UnaryOperation<DeviceType::Cuda, int, TOutput>

Constructor & Destructor Documentation

◆ CudaEncoderOp() [1/2]

template<typename TInput , typename TOutput = TInput>

Mila::Dnn::Compute::CudaEncoderOp< TInput, TOutput >::CudaEncoderOp ( const EncoderConfig & config )

inline

Constructs a new CUDA Encoder operation with the default device context.

Initializes the operation with a CUDA device context (defaults to CUDA:0).

◆ CudaEncoderOp() [2/2]

template<typename TInput , typename TOutput = TInput>

Mila::Dnn::Compute::CudaEncoderOp< TInput, TOutput >::CudaEncoderOp	(	std::shared_ptr< DeviceContext >	context,
		const EncoderConfig &	config
	)

inline

Constructs a new CUDA Encoder operation with a specific device context.

Parameters

context The device context to use for this operation.

Exceptions

std::runtime_error If the context is not for a CUDA device.

Member Function Documentation

◆ backward()

template<typename TInput , typename TOutput = TInput>

void Mila::Dnn::Compute::CudaEncoderOp< TInput, TOutput >::backward	(	const Tensor< int, MR > &	input,
		const Tensor< TOutput, MR > &	output,
		const Tensor< TOutput, MR > &	output_gradient,
		const std::vector< std::shared_ptr< Tensor< TOutput, MR > > > &	parameters,
		std::vector< std::shared_ptr< Tensor< TOutput, MR > > > &	parameter_gradients,
		Tensor< int, MR > &	input_gradient,
		const OperationAttributes &	properties,
		const std::vector< std::shared_ptr< Tensor< TOutput, MR > > > &	output_state
	)		const

inline

Performs the backward pass of the Encoder operation.

Computes gradients with respect to the embedding tables (token and position).

Parameters

input	Input tensor from the forward pass.
output	Output tensor from the forward pass.
output_gradient	Gradient of the loss with respect to the output.
parameters	Parameters tensor from forward pass.
parameter_gradients	Gradients for parameters (embedding tables).
input_gradient	Gradient of the loss with respect to the input (typically not used for discrete inputs).
properties	Additional attributes for the operation.
output_state	Cache tensors from forward pass.

Here is the call graph for this function:

◆ forward()

template<typename TInput , typename TOutput = TInput>

void Mila::Dnn::Compute::CudaEncoderOp< TInput, TOutput >::forward	(	const Tensor< int, MR > &	input,
		const std::vector< std::shared_ptr< Tensor< TOutput, MR > > > &	parameters,
		const OperationAttributes &	properties,
		Tensor< TOutput, MR > &	output,
		std::vector< std::shared_ptr< Tensor< TOutput, MR > > > &	output_state
	)		const

inlineoverride

Performs the forward pass of the Encoder operation on CUDA.

Transforms input token IDs into continuous embeddings by:

Looking up token embeddings from the embedding table (wte)
Adding positional embeddings (wpe) based on token position

The computation is performed on the GPU using CUDA kernels for optimal performance.

Parameters

input	Input tensor of shape [B, TDataType] containing token IDs, where B is batch size and TDataType is sequence length.
parameters	Vector of parameter tensors [wte, wpe] where wte is of shape [V, C] (vocabulary size � embedding dimension) and wpe is of shape [maxT, C] (maximum sequence length � embedding dimension).
properties	Additional attributes for the operation.
output	Output tensor of shape [B, TDataType, C] containing the resulting embeddings.
output_state	Cache for intermediate results (not used in this operation).

Here is the call graph for this function:

◆ getName()

template<typename TInput , typename TOutput = TInput>

std::string Mila::Dnn::Compute::CudaEncoderOp< TInput, TOutput >::getName ( ) const

inlineoverridevirtual

Gets the name of this operation.

Returns: std::string The name of the operation ("Cuda::EncoderOp").

Implements Mila::Dnn::Compute::OperationBase< TDeviceType, TInput1, TInput2, TOutput >.

Member Data Documentation

◆ config_

template<typename TInput , typename TOutput = TInput>

EncoderConfig Mila::Dnn::Compute::CudaEncoderOp< TInput, TOutput >::config_

private

Configuration for the encoder operation.

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

/home/runner/work/Mila/Mila/Mila/Src/Dnn/Compute/Operations/Cuda/CudaEncoderOp.ixx

Public Types

Public Member Functions

Private Attributes

Detailed Description

Member Typedef Documentation

◆ MR

◆ OperationBase

Constructor & Destructor Documentation

◆ CudaEncoderOp() [1/2]

◆ CudaEncoderOp() [2/2]

Member Function Documentation

◆ backward()

◆ forward()

◆ getName()

Member Data Documentation

◆ config_