Accelerating MoE with IO and Tile-aware Optimizations

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

TurboDiffusion: 100–200× Acceleration for Video Diffusion Models

Tile-Based Runtime for Ultra-Low-Latency LLM Inference

a size profiler for cuda binary

NVIDIA cuTile learn

cuTile is a programming model for writing parallel kernels for NVIDIA GPUs

Helpful kernel tutorials and examples for tile-based GPU programming

Code for the paper “Four Over Six: More Accurate NVFP4 Quantization with Adaptive Block Scaling”

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

Distributed MoE in a Single Kernel [NeurIPS '25]

Advanced quantization toolkit for LLMs and VLMs. Support for WOQ, MXFP4, NVFP4, GGUF, Adaptive Schemes and seamless integration with Transformers, vLLM, SGLang, and llm-compressor

Low overhead tracing library and trace visualizer for pipelined CUDA kernels

GPU programming related news and material links

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

Triton-based Symmetric Memory operators and examples

A framework to compare low-bit integer and float-point formats

StreamingVLM: Real-Time Understanding for Infinite Video Streams

Multi-Level Triton Runner supporting Python, IR, PTX, and cubin.

NVIDIA NVSHMEM is a parallel programming interface for NVIDIA GPUs based on OpenSHMEM. NVSHMEM can significantly reduce multi-process communication and coordination overheads by allowing programmer…

QuTLASS: CUTLASS-Powered Quantized BLAS for Deep Learning