gpu-kernels

memra is a Rust + CUDA inference engine built for NVIDIA RTX PRO 6000 Blackwell and RTX 5090. It serves GGUF models over an OpenAI-compatible API, and every default it ships was measured on those two cards — both Blackwell sm_120a, with a separately compile-gated Hopper/H100 (sm_90a) lane alongside.

[DEPRECATED] Moved to ROCm/rocm-systems repo

AgentKernelArena provides an end-to-end siloed-benchmarking environment where different LLM-powered agents—such as Cursor Agent, Claude Code, Codex, SWE-agent, and GEAK—can be evaluated side-by-side on the same GPU kernel tasks, using objective and reproducible metrics.

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

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

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

Source code and documentation for Flash-Aurora porject.

GPU Drano Static Analysis for GPU programs.

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

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

The official implementation for paper "AdaExplore: Failure-Driven Adaptation and Diversity-Preserving Search for Efficient Kernel Generation"

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

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

Apple Silicon (Metal) backend for OpenAI Triton: write standard @triton.jit kernels on your Mac GPU. The same source runs on NVIDIA (fp32 verified bit-identical), so you develop kernel logic locally and rent a GPU only for the perf pass.

Native Blackwell (sm_100) tcgen05 training backward for the gated-linear-recurrence family (GDN-2/GLA/KDA/SSD), plus a contract-grade verifier that falsifies published GPU kernels. Six open Mamba-3 Triton kernels including the first open MIMO backward; GRPO trainer graded on contract correctness. state-spaces/mamba#904

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.

Agent-queryable ROCm kernel optimization knowledge base for AMD Instinct MI300/gfx942 and MI350/MI355X/gfx950, packaged for Codex CLI and Claude Code with merged-PR provenance, real-silicon validation, and a maintainer-controlled pull-request evidence pipeline.

ORDA: Triton kernels for logit-level knowledge distillation loss.