IBM Spyre Accelerator Moves torch.compile Backend to Inductor
IBM's PyTorch software stack for its Spyre accelerator has switched its torch.compile backend from sendnn, a separate compilation stack built from scratch, to backend=inductor, the open-source backend. The change is detailed in a PyTorch blog post titled "Extending Inductor for the IBM Spyre Accelerator."
Sources and evidence
Summary last validated Oct 13, 2026
Reader actions
Report an issue
Use this for an incorrect summary, wrong source, duplicate story, or wrong category. Submissions are private and do not change the story automatically.
Related coverage
- open source
PyTorch Helion integrates NVIDIA CompileIQ for compiler controls
PyTorch's blog describes how Helion, its tool for writing GPU kernels in PyTorch-native Python with an autotuner, now incorporates NVIDIA CompileIQ to add compiler controls. The post frames this as a two-layer search approach, pairing Helion's autotuning with CompileIQ's compiler-level controls.
- infrastructure
PyTorch Introduces TorchNeuron Native Backend for AWS Trainium
PyTorch announced TorchNeuron, a native PyTorch backend for AWS Trainium. The blog post says frontier models now span language, video and audio across dense and mixture-of-experts transformers, diffusion, state-space, world and structure-prediction architectures, and frames PyTorch as where AI research and models are built. The post provides no further technical details in the excerpt.
- chips
Nvidia RTX Spark laptops launch from ASUS, Dell, HP, Lenovo, Microsoft and MSI starting at $2,399
Nvidia's RTX Spark laptops are now available, bringing Blackwell RTX graphics, up to 128GB of unified memory and the full Nvidia AI platform to slim Windows PCs. Models from ASUS, Dell, HP, Lenovo, Microsoft and MSI start at $2,399, targeting local AI, content creation and gaming on a portable device.
- chips
NVIDIA RTX Spark Runs Local AI Coding Agents on GB10
NVIDIA introduced RTX Spark, a system built for AI development that runs coding agents locally on one machine. It uses the NVIDIA GB10 system-on-a-chip and large unified memory, replacing the split workflow of laptop, cloud model API, and cloud VM. NVIDIA says developers can code, test, and deploy locally.
- open source
PyTorch Blog Introduces Zephon for Deterministic Data Loading in TorchTitan
A PyTorch blog post describes Zephon, a system for bringing elastically deterministic data loading to TorchTitan, PyTorch's large-scale training framework. The post frames training data as a key lever for model performance and says much of the day-to-day work involves managing it, but provides few technical details in the available excerpt.