gpu-kernels

GPU Profiler for AI agents

An agentic system that auto-optimizes LLM workloads on AMD GPUs.

[DEPRECATED] Moved to ROCm/rocm-systems repo

AgentKernelArena is a platform for evaluating and training AI agents for GPU kernel optimization, combining realistic tasks, correctness validation, and performance-based rewards for reinforcement learning.

Production-Grade Autoresearch. Ideal for agent harness engineering, prompt engineering, ML model development, GPU kernels, and other optimizable code.

Open source skill library for AI coding agents to write, optimize, and debug high performance compute kernels across CUDA, Triton, and quantized workloads.

Online CUDA Occupancy Calculator

Agents, and RL environment, for optimizing GPU kernels on AMD ROCm using LLM agents. Benchmarks LLM serving workloads end-to-end, profiles bottleneck kernels, optimizes them via Claude Code or Codex, and scores on compilation, correctness, and speedup.

(Spring 2017) Assignment 2: GPU Executor

Source code and documentation for Flash-Aurora porject.

Minimal FlashAttention in CUDA C++/CuTe: readable WMMA/CuTe kernels, no NxN workspace, up to 4.5x faster than naive PyTorch

Runs a single CUDA/OpenCL kernel, taking its source from a file and arguments from the command-line

GPU Drano Static Analysis for GPU programs.

Apple Silicon (Metal) backend for Triton: write standard @triton.jit kernels on your Mac GPU. The same source runs bit-identical on NVIDIA and AMD, so you develop kernel logic locally and rent a datacenter GPU only for the perf pass.

[COLM-LLA 2026] The official implementation for paper "AdaExplore: Failure-Driven Adaptation and Diversity-Preserving Search for Efficient Kernel Generation"

Run a coding model at 300+ tok/s on your Mac. Native Rust + Metal inference engine for LiquidAI LFM2.5-24B-A2B (MLX 4-bit) on Apple Silicon.

Prototype for a SPIR-V assembler and dissasembler. It provides a composable Java interface for generating SPIR-V code at runtime.

Native Blackwell (sm_100) tcgen05 training backward for the gated-linear-recurrence family (GDN/KDA/GLA/SSD/LA), plus a contract-grade verifier that audits LLM-generated GPU kernels. Six Mamba-3 Triton kernels; GRPO trainer graded on contract correctness. arXiv:2608.12700

Repo containing artifacts for Neurips 2025 tutorial- How to Build Agents to Generate Kernels for Faster LLMs (and Other Models!)

From-scratch reimplementation of DeepSeek's Native Sparse Attention (arXiv:2502.11089) in Triton + CUDA Hopper WGMMA. 7.07x faster than FlashAttention-3 at 64k context. Five-model training fleet, perplexity sweep, LongBench v2, MoBA comparison.