megatron.model.fused_softmax#

Description

Classes

FusedScaleMaskSoftmax(input_in_fp16, ...)

fused operation: scaling + mask + softmax

ScaledMaskedSoftmax(*args, **kwargs)

Fused operation which performs following three operations in sequence 1.

ScaledSoftmax(*args, **kwargs)

Fused operation which performs following two operations in sequence 1.

ScaledUpperTriangMaskedSoftmax(*args, **kwargs)

Fused operation which performs following three operations in sequence 1.