IBM Spyre becomes a first-class PyTorch backend
IBM connected its Spyre accelerator to the PyTorch ecosystem as a first-class backend through torch-spyre, a vLLM plugin and compiler updates.
AT A GLANCE
Key points
- torch-spyre compiles ordinary PyTorch code into an execution plan for Spyre.
- IBM says its new Inductor-based compiler contains roughly seven times less custom code.
- The Spyre vLLM plugin path has been validated with Gemma 4, Qwen3 and Granite 4.2.
What is changing?
IBM Research announced software updates that make its dataflow-based Spyre AI accelerator a first-class PyTorch backend. At the center, torch-spyre compiles ordinary PyTorch code into an execution plan for Spyre. IBM aims to let developers use models from ecosystems such as Hugging Face on IBM Z, LinuxONE and Power systems through more familiar workflows.
The update covers an Inductor-based compiler path, a vLLM plugin for Spyre, hardware-agnostic model layers and continuous-integration testing. IBM says the new compiler contains roughly seven times less custom code than the system it replaced. The company also reports 1.7-times faster prefill and 2.4-times faster decode in a specific Granite 3.3 8B test; those results apply to the disclosed configuration.
The vLLM plugin pathway has been validated with Gemma 4, Qwen3 and Granite 4.2. Spyre’s PyTorch backend also provides the foundation for connecting the accelerator to the distributed-inference layer llm-d.
Explore the original sources
Source date: 26 October 2026.
Published by AI Radar: 26 Oct 2026, 18:06 (Istanbul, UTC+3). Updated: 26 Oct 2026, 18:09.
FOLLOW THE TOPIC
Topics and tags
Share your discovery.
Send this story to a friend.
Reader comments
What do you think about this development? Share your experience or question. Comments are published after editorial approval.
Your name/nickname and comment become public if approved. Read: Privacy notice.
Loading comments…