Pinned Loading

1. MoE-Kernel-Agent MoE-Kernel-Agent Public

   🥈 2nd Place (Full-Agent) · MLSys 2026 FlashInfer AI Kernel Generation Contest — MoE Track (NVIDIA) · 25.7× median speedup over PyTorch reference

   Python 8 1

2. Triton-Attention-Kernels Triton-Attention-Kernels Public

   GPU attention kernels for LLM inference, implemented from scratch in Triton.

   Jupyter Notebook

3. DeepSeek-V3-FP8-MoE-Kernel DeepSeek-V3-FP8-MoE-Kernel Public

   High-performance fused MoE kernel for DeepSeek-V3 FP8 inference on Blackwell (B200), written in Triton. 1.08× faster than FlashInfer on total workload latency.

   Python

4. SVD-Flash SVD-Flash Public

   SVD-Flash: Efficient LLM inference via SVD Compression and Tiling on AWS Trainium

   Python

5. ThunderAgent ThunderAgent Public

   Forked from ThunderAgent-org/ThunderAgent

   A simple, fast and robust program-aware agentic inference system.

   HTML

6. nki-llama nki-llama Public

   Python