Efficiency: DPad-enhanced dLLMs achieve up to a 61.39× speedup over vanilla dLLM baselines. Accuracy: DPad-enhanced dLLMs achieve up to a +26.46% improvement over vanilla dLLM baselines. (Evaluation conducted on NVIDIA A100-PCIe-80GB GPUs).
Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models
DMax is a new dLLM paradigm achieving aggressive parallel decoding while preserving generation quality.
Code accompanying the paper "Layer Collapse in Diffusion Language Models" by Alexander Conzelmann, Albert Catalan-Tatjer, and Shiwei Liu (Tübingen AI Center / MPI for Intelligent Systems / ELLIS Institute Tübingen). Link: https://arxiv.org/abs/2605.06366
The pipeline of dKV-Cache
Die Hauptfunktionen von horseee/dkv-cache sind: Inference Optimization.
Open-Source-Alternativen zu horseee/dkv-cache sind unter anderem: crys-chen/dpad — Efficiency: DPad-enhanced dLLMs achieve up to a 61.39× speedup over vanilla dLLM baselines. Accuracy: DPad-enhanced… cychomatica/freedave — Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models. czg1225/dmax — DMax is a new dLLM paradigm achieving aggressive parallel decoding while preserving generation quality. danielmisrael/apd — Official repository for the paper: Accelerating Diffusion LLMs via Adaptive Parallel Decoding. duterscmy/soar — . conzel/super-outlier-dlm — Code accompanying the paper "Layer Collapse in Diffusion Language Models" by Alexander Conzelmann, Albert…