stateful-models

SpecLA proposes a speculative decoding runtime tailored for stateful linear-attention models, achieving up to 1.70x end-to-end speedup over autoregressive decoding on an NVIDIA H100 with a GDN-1.3B target.