AWS Neuron SDK Release Notes - v2.28.0

Release Date: February 25, 2026

---

Today we are releasing AWS Neuron SDK 2.28.0. This release enhances Neuron Explorer with system profiling, Tensor Viewer, and Database Viewer for comprehensive performance analysis. NxD Inference adds support for Qwen2/Qwen3 VL vision language models, Flux.1 inpainting capabilities, and Eagle3 speculative decoding. The NKI Library expands with 9 new kernels including RoPE, MoE operations, and experimental kernels for attention and cross entropy. NKI (Beta 2) introduces LNC multi-core support with intra-LNC collectives and new APIs. Kubernetes users gain Neuron DRA Driver support for advanced resource allocation.

---

Developer Tools and Profiling

Neuron Explorer Enhancements — Added system profiling support with drill-down navigation to device profiles. New Tensor Viewer helps identify memory bottlenecks by displaying tensor names, shapes, sizes, and memory usage. Database Viewer provides an interactive interface for querying profiling data using SQL or natural language. Profile Manager now supports tag-based organization and search. A migration guide from Neuron Profiler/Profiler 2.0 is now available.

nccom-test Improvements — Enhanced data integrity checks use pseudo-random data patterns for better corruption detection. Added support for alltoallv collective operation for benchmarking variable-sized all-to-all communication patterns.

---

Inference Updates

NxD Inference 0.8.16251 — Added support for vision language models including Qwen2 VL (Qwen2-VL-7B-Instruct) and Qwen3 VL (Qwen3-VL-8B-Thinking) for processing text and image inputs (Beta). Pixtral model support improved with batch size 32 and sequence length 10240 on Trn2 with vLLM V1. Flux.1 model gains new functionality for in-paint, out-paint, canny edge detection, and depth-based image generation (Beta).

vLLM Neuron Plugin 0.4.0 — Multi-LoRA serving enhancements enable streaming LoRA adapters via vLLM's load_adapter API with dynamic runtime loading. Users can now run the base model alone when multi-LoRA serving is enabled. Added Eagle3 speculative decoding support for Llama 3.1 8B. Updated to support vLLM v0.13.0 and PyTorch 2.9.

---

NKI Library

9 New Kernels — The NKI Library expands from 7 to 16 documented kernel APIs. New core kernels include:

• RoPE — Rotary Position Embedding
• Router Top-K — Expert selection for MoE
• MoE CTE — Context Encoding
• MoE TKG — Token Generation
• Cumsum — Cumulative sum

New experimental kernels include:

• Attention Block TKG — Fused attention for token generation
• Cross Entropy — Forward and backward passes
• Depthwise Conv1D
• Blockwise MM Backward — For MoE training

Enhanced Quantization Support — Existing kernels receive FP8 and MX quantization support across QKV, MLP, and Output Projection kernels. QKV kernel adds fused FP8 KV cache quantization and block-based KV cache layout. MLP kernel adds gate/up projection clamping and fp16 support for TKG mode. Attention CTE kernel adds strided Q slicing for context parallelism.

Improved Utilities — TensorView gains rearrange method for dimension reordering and has_dynamic_access for runtime-dependent addressing checks. SbufManager provides hierarchical tree-formatted allocation logging with new query methods for SBUF utilization. New utilities include rmsnorm_mx_quantize_tkg, interleave_copy, LncSubscriptable, and TreeLogger.

---

Neuron Kernel Interface (NKI)

NKI Beta 2 (0.2.0) — This release includes LNC multi-core support for LNC=2, enabling kernels to leverage multiple NeuronCores within a logical NeuronCore. The compiler now tracks shared_hbm tensors and canonicalizes LNC kernel outputs. Users can declare tensors private to a single NeuronCore using private_hbm memory type.

New nki.collectives Module — Enables collective communication across multiple NeuronCores with operations including:

• all_reduce
• all_gather
• reduce_scatter
• all_to_all
• collective_permute variants
• rank_id

New APIs and Features — New nki.isa APIs include nonzero_with_count for sparse computation and exponential for element-wise operations. New float8_e4m3fn dtype supports FP8 workloads. Language features include no_reorder blocks for instruction ordering control, __call__ special method support, tensor.view method for reshaping, and shared constants as string arguments.

API Improvements — dma_transpose now supports indirect addressing, dma_copy adds the unique_indices parameter, and register_alloc accepts optional tensor arguments for pre-filling. The compiler no longer truncates diagnostic output.

---

Kubernetes Support

Neuron DRA Driver — Introduced Neuron Dynamic Resource Allocation (DRA) Driver enabling advanced resource allocation using the Kubernetes DRA API for flexible and efficient Neuron device management. The DRA API provides topology-aware scheduling, atomic resource allocation, and per-workload configuration. Neuron Helm Charts now include DRA Driver support.

---

PyTorch Framework (torch-neuronx)

Transition to Native PyTorch Support — Starting with PyTorch 2.10 support (planned for a future Neuron release), AWS Neuron will transition from PyTorch/XLA to native PyTorch support via TorchNeuron. PyTorch 2.9 is the last version using PyTorch/XLA. Users will need to update their scripts when upgrading to PyTorch 2.10 or later.

---

For the full component-level release notes, see the Neuron 2.28.0 Component Release Notes.