inference-acceleration

[ICLR2025, ICML2025, NeurIPS2025 Spotlight] Quantized Attention achieves speedup of 2-5x compared to FlashAttention, without losing end-to-end metrics across language, image, and video models.

TurboDiffusion: 100–200× Acceleration for Video Diffusion Models

Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model

[ICML2025] SpargeAttention: A training-free sparse attention that accelerates any model inference.

SLA: Beyond Sparsity in Diffusion Transformers via Fine-Tunable Sparse–Linear Attention

SparkDiffusion: Accelerating DiT video generation by 265× via joint Sparsity, Distillation, Quantization, etc.

Less is Enough: Training-Free Video Diffusion Acceleration via Runtime-Adaptive Caching

A unified, torch-native training framework for MTP and block-parallel speculative decoding.

[ICLR 2026] Discrete Diffusion Forcing (D2F): dLLMs Can Do Faster-Than-AR Inference

[NeurIPS 2024] AsyncDiff: Parallelizing Diffusion Models by Asynchronous Denoising

[EMNLP 2026] 📚 A curated list of Awesome Efficient dLLMs Papers with Codes

⚡️ A fast and flexible PyTorch inference server that runs locally, on any cloud or AI HW.

KsanaDiT: High-Performance DiT (Diffusion Transformer) Inference Framework for Video & Image Generation

This is the official repo of "QuickLLaMA: Query-aware Inference Acceleration for Large Language Models"

[CVPR 2026] VDE: Training-Free Accelerating Rectified Flow Model via Velocity Decomposition and Estimation

The source code of "Merging Experts into One: Improving Computational Efficiency of Mixture of Experts (EMNLP 2023)":

An operating system for LLM inference that makes the most of your hardware: generate optimized compute, run models beyond available memory, and orchestrate and migrate layers across CPUs and GPUs at runtime to improve throughput without losing the conversation.

[ICML26] Official Repo for WorldCache: Accelerating World Models for Free via Heterogeneous Token Caching

Mixed-precision acceleration patch for MiniMax H3 inference on NVIDIA V100

[ICLR 2026] Learning to Parallel: Accelerating Diffusion Large Language Models via Learnable Parallel Decoding