Popular repositories Loading

1. cuda-optimized-skill cuda-optimized-skill Public

   A CUDA kernel optimization toolkit for validation, benchmarking, Nsight Compute profiling, bottleneck analysis, and iterative tuning. It helps improve custom GPU operators with reproducible workflo…

   Python 214 19

2. cuda-opt-agent cuda-opt-agent Public

   CUDA Opt Agent is an LLM-driven toolkit for automated CUDA kernel optimization. It generates baselines from text, specs, or existing .cu files, compiles and validates with nvcc, benchmarks on real …

   Python 14 3

3. cuda-datagen cuda-datagen Public

   Multi-agent LangGraph pipeline that turns CUDA kernel questions into verified SFT data: LLM-generated CUDA/CUTLASS/Triton/TileLang kernels pass nvcc + real GPU numeric checks, critic review and ran…

   Python 1

4. cuda-sft-qgen cuda-sft-qgen Public

   cuda-sft-qgen: a LangGraph multi-agent pipeline generating CUDA SFT questions from a versioned knowledge graph. Every question passes 3-layer dedup, independent verification and a separate reviewer…

   Python

5. lightly-trian-quant lightly-trian-quant Public

   RTX 5090 optimization of DINOv3 + EoMT segmentation training and inference with fused CUDA kernels, CUDA Graphs, and FP8/FP4 quantization. Achieves 6.3x ViT-L inference and 4.22x training speedups …

   Python

6. FlashMoE FlashMoE Public

   Forked from osayamenja/FlashMoE

   Fork of osayamenja/FlashMoE, the fused distributed MoE kernel, with consumer Blackwell SM120/SM121 (RTX 50 / RTX PRO 6000 / DGX Spark) support and MXFP8, MXFP4 and NVFP4 expert paths: up to 5.6x fa…

   Cuda