DFloat11 [NeurIPS '25]: Lossless Compression of LLMs and DiTs for Efficient GPU Inference

A library of GPU kernels for sparse matrix operations.

SpInfer: Leveraging Low-Level Sparsity for Efficient Large Language Model Inference on GPUs

Helpful kernel tutorials and examples for tile-based GPU programming

⚡FlashRAG: A Python Toolkit for Efficient RAG Research (WWW2025 Resource)

Accelerating MoE with IO and Tile-aware Optimizations

[ASPLOS'26] Taming the Long-Tail: Efficient Reasoning RL Training with Adaptive Drafter

The official repository for PTQTP implementation

Trainable fast and memory-efficient sparse attention

🚀🚀 Efficient implementations of Native Sparse Attention

PyTorch bindings for CUTLASS grouped GEMM.

A tiny yet powerful LLM inference system tailored for researching purpose. vLLM-equivalent performance with only 2k lines of code (2% of vLLM).

Official PyTorch Implementation of "Diffusion Transformers with Representation Autoencoders"

QeRL enables RL for 32B LLMs on a single H100 GPU.

StreamingVLM: Real-Time Understanding for Infinite Video Streams

A framework for serving and evaluating LLM routers - save LLM costs without compromising quality

Modeling, training, eval, and inference code for OLMo

Why Low-Precision Transformer Training Fails: An Analysis on Flash Attention

[CVPR 2025 Oral] Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models

Quantized LLM training in pure CUDA/C++.

A lightweight Inference Engine built for block diffusion models

VolSplat: Rethinking Feed-Forward 3D Gaussian Splatting with Voxel-Aligned Prediction

Implementation for FP8/INT8 Rollout for RL training without performence drop.

Paper reading and discussion notes, covering AI frameworks, distributed systems, cluster management, etc.

Research prototype of PRISM — a cost-efficient multi-LLM serving system with flexible time- and space-based GPU sharing.