torchspec

EAGLE 3.1 improves speculative decoding robustness with post-norm architecture, achieving up to 2x longer acceptance length in long-context workloads, with training support from TorchSpec and integration into vLLM.