Capability interface for position-dependent paired operations. More...

Inheritance diagram for Mila::Dnn::Compute::IPositionalPairedOp:

[legend]

Public Member Functions
virtual	~IPositionalPairedOp ()=default
virtual void	decode (const ITensor &inputA, const ITensor &inputB, ITensor &outputA, ITensor &outputB, int position)=0
	Process a single token at an explicit sequence position.
virtual void	prefill (const ITensor &inputA, const ITensor &inputB, ITensor &outputA, ITensor &outputB, int position_offset)=0
	Process a chunk of tokens starting at a given position.

Detailed Description

Capability interface for position-dependent paired operations.

Implemented by operations that accept two input/output tensor pairs and whose mathematical output changes based on absolute token position � e.g. RoPE, which rotates both Q and K embeddings by position-dependent angles.

Parameter order follows the PairedOperation convention: all inputs first, then all outputs, followed by the position argument.

Operations that are position-agnostic do not implement this interface � they use forward() for all modes.

Constructor & Destructor Documentation

◆ ~IPositionalPairedOp()

virtual Mila::Dnn::Compute::IPositionalPairedOp::~IPositionalPairedOp ( )

virtualdefault

Member Function Documentation

◆ decode()

virtual void Mila::Dnn::Compute::IPositionalPairedOp::decode	(	const ITensor &	inputA,
		const ITensor &	inputB,
		ITensor &	outputA,
		ITensor &	outputB,
		int	position )

pure virtual

Process a single token at an explicit sequence position.

Parameters

inputA	Single-token first input [B, 1, ...].
inputB	Single-token second input [B, 1, ...].
outputA	Single-token first output [B, 1, ...].
outputB	Single-token second output [B, 1, ...].
position	Zero-based absolute sequence position.

Implemented in Mila::Dnn::Compute::Cuda::Rope::CudaRopeOp< TComputePrecision >, Mila::Dnn::Compute::Cuda::Rope::CudaRopeOp< TensorDataType::BF16 >, and Mila::Dnn::Compute::Cuda::Rope::CudaRopeOp< TensorDataType::FP32 >.

◆ prefill()

virtual void Mila::Dnn::Compute::IPositionalPairedOp::prefill	(	const ITensor &	inputA,
		const ITensor &	inputB,
		ITensor &	outputA,
		ITensor &	outputB,
		int	position_offset )