A training-free inference framework for diffusion language models that executes multiple block-size branches in a single batched forward pass, reducing denoising NFEs by 26.6% and achieving 1.33× speedup over Fast-dLLM.
A co-design framework for serving hybrid Mamba-Transformer models on GPU–PIM heterogeneous hardware, achieving 1.51× throughput over prior art and 2.47× over GPU-only baselines while reducing tail latency to 0.61×.
An algorithm–system co-design that exploits cross-step temporal redundancy at three independent granularities — layers, tokens, and columns — achieving 2.66× and 2.30× throughput over dense baselines on A100 and H100 with negligible accuracy loss.