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

AWS Neuron SDK 2.26.0: Developer Tools release notes#

Date of release: September 18, 2025

Improvements #

The Neuron Profiler UI now allows you to select multiple semaphore values to display simultaneously for a more comprehensive view of activity.

nccom-test support on Trn2 for State Buffer to State Buffer collectives benchmarking for all-reduce, all-gather, and reduce-scatter operations.

System profile grouping default in Perfetto now uses global NeuronCore ID instead of process local NeuronCore ID for better display of multi-process workloads.
Added warning when system profile events are dropped due to limited buffer space, and added suggestion of how configure more buffer space if desired.
nccom-test will show helpful error message when invalid sizes are used with all-to-all collectives.

Here’s what we fixed in 2.26.0:

Fixed device memory usage type table and improvement made to stay in sync between runtime and tools versions.
Fixed system profile crash when processing long-running workloads.
Fixed display of system profiles in Perfetto to correctly separate rows within the same Logical NeuronCore when using NEURON_LOGICAL_NC_CONFIG=2 on Trn2.

System profile hardware events may be misaligned due to sync point imprecision. In Perfetto, this may cause events to be interleaved.
System profile events shown in the Neuron Profiler UI for multiprocess workloads are grouped together. Please try the Perfetto output if you encounter this issue.
Currently, only a Neuron Runtime trace can be shown when capturing a system profile for a PyTorch workload. (Full framework traces can be shown for JAX workloads, though.) We are working to bring PyTorch traces into parity in a future release.

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