Maps abstract TensorDataType to host-compatible C++ type and TensorDataType. More...

Detailed Description

template<TensorDataType TDataType>
struct Mila::Dnn::TensorHostTypeMap< TDataType >

Maps abstract TensorDataType to host-compatible C++ type and TensorDataType.

Defines the preferred host-side type and corresponding TensorDataType for each abstract tensor data type, used for host-device conversions, fill operations, and host-side processing. This mapping preserves type characteristics where possible while ensuring host compatibility.

Design decisions:

Floating-point device types (FP16, BF16, FP8) map to float for host compatibility
Integer types map to their corresponding fixed-width standard types
Maintains precision and signedness characteristics of original types
Provides direct host representation for device-agnostic operations

Template Parameters

TDataType Abstract tensor data type from TensorDataType enumeration

Note: This is a closed template - only explicit specializations are valid; Host types are device-agnostic and consistent across all compute backends; Instantiating the primary template produces a clear compile-time error

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

/__w/Mila/Mila/Mila/Src/Dnn/Tensors/TensorHostTypeMap.ixx