IBM says Spyre has become a first-class PyTorch backend

Oct 26

1. 20:29

   Unite.AI

   IBM Makes Its Spyre Accelerator a First-Class PyTorch Backend

   IBM announced that software updates in the PyTorch ecosystem in the second half of 2026 have made the Spyre dataflow AI accelerator a first-class PyTorch backend. The torch.compile backend of torch-spyre has migrated from the in-house sendnn to an Inductor-based extension, reducing the codebase from about 273,000 lines to about 38,600 lines, a roughly 7x reduction. In a demo with Granite 3.3 8B, batch size 1, and sequence length 1024, prefill was 1.7x faster and decode was 2.4x faster.

   Expand summary