Mila::Dnn::Compute::IPositionalDecode Struct Reference

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

Inheritance diagram for Mila::Dnn::Compute::IPositionalDecode:

Public Member Functions
virtual	~IPositionalDecode ()=default
virtual void	decode (const ITensor &input, ITensor &output, int position)=0
	Process a single token at an explicit sequence position.

Detailed Description

Capability interface for position-dependent unary operations.

Implemented by operations whose mathematical output changes based on the token's absolute position in the sequence — e.g. positional embedding lookups (Lpe) and attention operations that index into a KV cache (GQA, MHA).

Operations that are position-agnostic (Linear, RmsNorm, SwiGLU, Residual) do not implement this interface — they use forward() for all modes.

Constructor & Destructor Documentation

~IPositionalDecode()

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

virtualdefault

Member Function Documentation

decode()

virtual void Mila::Dnn::Compute::IPositionalDecode::decode ( const ITensor & input, ITensor & output, int position )

pure virtual

Process a single token at an explicit sequence position.

Parameters

input	Single-token input [B, 1, ...].
output	Single-token output [B, 1, ...].
position	Zero-based absolute sequence position.

Implemented in Mila::Dnn::Compute::Cuda::Lpe::CudaLpeOp< TInput, TPrecision >, Mila::Dnn::Compute::Cuda::Lpe::CudaLpeOp< TensorDataType::INT32, TensorDataType::BF16 >, and Mila::Dnn::Compute::Cuda::Lpe::CudaLpeOp< TensorDataType::INT32, TensorDataType::FP32 >.

---

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

• /__w/Mila/Mila/Mila/Src/Dnn/Compute/Operations/IPositionalDecode.ixx