This document is relevant for: Inf1, Inf2, Trn1, Trn2, Trn3

AWS Neuron SDK 2.26.0: Neuron Kernel Interface (NKI) release notes#

Date of release: September 18, 2025

Improvements#

New nki.language APIs#

  • nki.language.gelu_apprx_sigmoid - Gaussian Error Linear Unit activation function with sigmoid approximation.

Updated nki.language APIs#

  • nki.language.tile_size.total_available_sbuf_size constant - Added a new field, total_available_sbuf_size, that contains the returned total available SBUF size.

New nki.isa APIs#

  • nki.isa.select_reduce - Selectively copy elements with maximum reduction.

  • nki.isa.sequence_bounds - Compute sequence bounds of segment IDs.

  • nki.isa.dma_transpose - Enhanced with:

    • axes parameter to define 4D transpose for supported cases

    • dge_mode parameter to specify Descriptor Generation Engine (DGE)

  • nki.isa.activation - Supports the new nl.gelu_apprx_sigmoid nki.language operation.

Improvements and fixes#

  • nki.language.store() - Supports PSUM buffer with extra additional copy inserted.

Documentation and tutorial updates#

  • Added documentation and example for nki.isa.dma_transpose API

  • Improved nki.simulate_kernel example

  • Updated tutorial code to use nl.fp32.min instead of a magic number

Previous release notes#

This document is relevant for: Inf1, Inf2, Trn1, Trn2, Trn3