IBM reports broader PyTorch and vLLM support for Spyre accelerators

IBM says software updates expand support for its Spyre inference accelerator through a PyTorch backend, a vLLM plugin, compiler improvements, and model testing. The integration lets ordinary PyTorch code compile for Spyre and provides a path for llm-d to schedule inference work onto the accelerator. IBM says early hardware-agnostic support in vLLM has been validated with the Spyre plugin on Gemma 4, Qwen3, and Granite 4.2. In a separate Granite 3.3 8B demonstration at batch size 1 and sequence length 1024, moving graph construction out of data transfer and compute dispatch produced 1.7 times faster prefill and 2.4 times faster decode. These performance figures describe that specific demonstration.