Release Date: August 17, 2026
Today we are releasing Neuron 2.32.0. This release includes updates across the stack: NKI 0.6.0 — on-device top-K reduction, variable-length collectives for ragged per-rank data, and data-dependent runtime loops; 13 new NKI Library kernels for DeepSeek-V3.2 sparse-MLA context encoding and MXFP8 Mixture-of-Experts training; explicit native 64-bit integer control in the Neuron Compiler; uneven per-rank data distribution via variable-size collectives in the Runtime on Trn2 and Trn3; and a new Neuron Agentic Development skill that ports HuggingFace models to the vLLM Neuron backend. This release also upgrades vLLM Neuron, launched in July 2026, to vLLM 0.24.0.
What’s in this release
-
NKI 0.6.0 — Adds a GpSimd Engine top-K instruction via nisa.topk, variable-length collectives for ragged per-rank data via all_gather_v, and new runtime-loop constructs (fori_loop/while_loop) for data-dependent iteration that replace nl.dynamic_range. Relaxes DMA transpose constraints. See Neuron Kernel Interface (NKI) [0.6.0] (Neuron 2.32.0 Release).
-
NKI Library — Adds 12 new experimental kernels and 1 new core kernel, including the split DeepSeek-V3.2 sparse-MLA context path, MXFP8 flash-decode attention, MXFP8 blockwise-MoE forward paired with an MXFP8 matmul backward for training, a fused GPT-OSS sliding-window-attention block, and a GpSIMD top-K. Adds PyTorch reference implementations for 22 more kernels. See NKI Library (NKI-Lib) (Neuron 2.32.0 Release).
-
Graph Compiler — Adds explicit control over 64-bit integer compilation via --native-int64 and --implicit-integer-downcast, expands complex64 coverage to 30 operations, and rewrites one-hot embedding lookups into gather operations, reducing compile time up to 64% and NEFF size up to 96% for affected patterns. Adds MXFP8 factorized scale layout and reduced weight-load frequency on Trn3. See Neuron Compiler [2.27.5334.0] (Neuron 2.32.0 Release).
-
Runtime & Driver — Adds variable-size collectives (AllGatherV, ReduceScatterV, AllToAllV) for workloads with uneven per-rank data on Trn2 and Trn3, one-rank-per-die ring and barrier support on the Trn3 Gen2 UltraServer topology, and the Mesh collective algorithm in multi-stream mode. Raises the maximum NCCL communicators per NEFF from 12 to 16. See Neuron Runtime (Neuron 2.32.0 Release).
-
Neuron Agentic Development — A new neuron-framework-autoport-vllm-neuron skill ports HuggingFace models to the vLLM Neuron backend end to end, and the model-equivalence skill now validates vLLM Neuron ports. Updated for NKI 0.6.0. See Neuron Agentic Development (Neuron 2.32.0 Release).
-
Neuron Explorer — Adds per-core host CPU utilization to the System Trace Viewer, CSV export of the current timeline selection from the Device Trace Viewer, a Tensor Engine throttling explainer with an automatic small-DMA-transfer warning, and profile folders and deletion in the Profile Manager. The neuron-explorer capture CLI replaces neuron-profile capture. See Neuron Developer Tools & Neuron Explorer (Neuron 2.32.0 Release).
-
vLLM Neuron — Upgraded to vLLM 0.24.0 (vllm-neuron 0.24.0.1.1.0) and integrated into the 2.32.0 DLAMIs and DLCs. See vLLM Neuron Release Notes.
-
DLAMIs & Containers — All packages upgraded to Neuron SDK 2.32.0, with JAX 0.10.0 and vLLM 0.24.0. See Neuron DLAMIs (Neuron 2.32.0 Release) and Neuron Containers (Neuron 2.32.0 Release).