Topics / #cuda-kernels

Cuda-kernels projects on GitHub

Custom CUDA kernel development for GPU acceleration and performance study.

As of , the most-starred Cuda-kernels projects on GitHub are flash-attention (25.2K stars), cupy (12.4K stars) and taskflow (12.2K stars). GitHub.Locker tracks 9 Cuda-kernels repositories with 81.5K stars combined. The fastest-growing on October 17 was VisionHOPE, with +98 stars in a day.

9Projects
81.5KStars in total
+206Stars in a day
Browse in the catalog →github.com/topics/cuda-kernels ↗

Hot right now

Most stars gained in a day · Oct 17

+98

VisionHOPE

PSRben/VisionHOPE · ★ 1.3K

VisionHOPE is the official PyTorch implementation of a visual backbone built on self-referential nested learning (SRNL), supporting ImageNet-1K classification, COCO detection/instance segmentation, and ADE20K semantic segmentation.

+86

StartLux-Decision

StartLuxLabs/StartLux-Decision · ★ 648

StartLux-Decision is a family of typed decision models (five dense sizes from 0.8B to 27B plus a 35B-A3B mixture-of-experts) that answer a question with a probability for each option instead of generating text.

+7

tilelang

tile-ai/tilelang · ★ 8.6K

TileLang (tile-lang) is a domain-specific language for writing high-performance GPU/CPU/NPU kernels like GEMM, Dequant GEMM, FlashAttention, and LinearAttention, using Pythonic syntax backed by a TVM-based compiler.

+6

flash-attention

Dao-AILab/flash-attention · ★ 25.2K

The official implementation of FlashAttention and FlashAttention-2 — GPU kernels that compute exact attention faster and with less memory than standard PyTorch attention.

The all-time top

9 projects · by stars

  1. #1 · Dao-AILab/flash-attention

    flash-attention

    The official implementation of FlashAttention and FlashAttention-2 — GPU kernels that compute exact attention faster and with less memory than standard PyTorch attention.

    25.2K
  2. #2 · cupy/cupy

    cupy

    CuPy is a NumPy/SciPy-compatible array library for GPU-accelerated computing with Python: you write ordinary Python array code, and it executes on NVIDIA or AMD GPUs.

    12.4K
  3. #3 · taskflow/taskflow

    taskflow

    Taskflow is a C++ library that helps you write high-performance task-parallel programs with high programming productivity, letting you build and run graphs of dependent tasks across CPU threads and GPUs.

    12.2K
  1. 4

    LeetCUDA

    12.1Kxlite-dev/LeetCUDA · Cuda

    LeetCUDA is an open-source book of CUDA learning notes for beginners, paired with a library of 200+ hand-written CUDA kernels (elementwise ops, HGEMM, Flash-Attention, etc.) and 100+ CUDA/LLM blog posts.

  2. 5

    DeepGEMM

    9Kdeepseek-ai/DeepGEMM · Cuda

    DeepGEMM is a unified CUDA kernel library that provides high-performance tensor core GEMM operations (FP8, FP4, BF16), fused MoE kernels, and attention-indexer kernels used in large language models such as DeepSeek.

  3. 6

    tilelang

    8.6Ktile-ai/tilelang · Python

    TileLang (tile-lang) is a domain-specific language for writing high-performance GPU/CPU/NPU kernels like GEMM, Dequant GEMM, FlashAttention, and LinearAttention, using Pythonic syntax backed by a TVM-based compiler.

  4. 7

    VisionHOPE

    1.3KPSRben/VisionHOPE · Python

    VisionHOPE is the official PyTorch implementation of a visual backbone built on self-referential nested learning (SRNL), supporting ImageNet-1K classification, COCO detection/instance segmentation, and ADE20K semantic segmentation.

  5. 8

    StartLux-Decision

    648StartLuxLabs/StartLux-Decision · Python

    StartLux-Decision is a family of typed decision models (five dense sizes from 0.8B to 27B plus a 35B-A3B mixture-of-experts) that answer a question with a probability for each option instead of generating text.

  6. 9

    v100-skinny

    174dnv2003/v100-skinny · Python

    v100-skinny provides hand-written NVFP4 (W4A16) and FP8 (W8A16) CUDA kernels that let four Tesla V100-SXM2-16GB GPUs run Qwen3.8-27B's published mixed NVFP4/FP8 weights directly, serving the model through an OpenAI-compatible API.

Often tagged together