Show HN: Auto GPU Kernel – Autonomous GPU-kernel discovery and optimizer

A new open-source tool, Auto GPU Kernel, autonomously discovers and optimizes GPU kernels, achieving a 34.93x average speedup to rank first in the DeepSeek Sparse Attention track of the MLSys 2026 FlashInfer AI Kernel Generation Contest. The tool uses an agentic loop with Claude Code CLI to iteratively profile, research, and optimize kernels, logging each experiment and adapting its optimization frequency as improvements diminish.

Autonomous GPU-kernel discovery & optimizer. Ranked 1 on MLSys 2026 - FlashInfer AI Kernel Generation Contest https://mlsys26.flashinfer.ai/ for the DeepSeek Sparse Attention DSA track with an average speedup of 34.93x. Submissions can be found at: | Kernel | Runtime ms | |---|---| | dsa topk indexer fp8 h64 d128 topk2048 ps64 /Dogacel/auto-gpu-kernel/blob/main/dsa topk indexer fp8 h64 d128 topk2048 ps64 — DSA TopK IndexerCopy the template directory into a separate folder / git repository to make sure your agents work in an isolated environment. The kernel agent is compatible with FlashInfer https://github.com/flashinfer-ai/flashinfer format and can run without a local GPU on cloud using Modal https://modal.com/ . Requires Claude Code CLI https://use-claude.com/index . Python env conda create -n fi-bench python=3.12 conda activate fi-bench pip install flashinfer-bench modal One-time environment setup modal setup modal volume create flashinfer-trace modal volume put flashinfer-trace /path/to/flashinfer-trace/ To get started clone the MLSys-2026 Contest Dataset https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest . To change the kernel you are implementing, please refer to the FlashInfer-Trace - Bring Your Own Kernel https://bench.flashinfer.ai/docs/tutorials/bring-your-own-kernel guide. Important Make sure you update CLAUDE.md to describe the kernel you are optimizing. The example in template is customized for sparse attention. Also optimize.md and benchmark.md has some parameters tuned for sparse attention such as number of test cases to run to get a sanity check. You can ask an agent to help you adjsut them. To run one iteration, claude --dangerously-skip-permissions -p "/optimize" Or you can launch interactive mode by running claude --dangerously-skip-permissions , selecting the right model, thinking mode and enter /loop Run /optimize every 15 minutes . That's it. The loop runs indefinitely, each iteration picks one optimization, benchmarks it, logs an experiment folder, and continues. Stop with Ctrl+C when you want to step in. As agent struggles to find new optimizations, it will start to change its schedule to be less frequent. For more details on the agentic loop, please refer to the technical report. Agents: - Profiler - Research - Workload inspector | Command | Purpose | |---|---| /optimize | Main loop | /benchmark <quick|stride N|full | One-shot Modal run | /log-experiment | Snapshot + write result.md + update index | See CLAUDE.md for rules and .claude/commands/ for full command specs. solution/triton/sparse fused.py — the kernel being optimized overwritten each iteration experiments/exp N/ — snapshot + results for iteration N experiments/summary.md — master index, one row per iteration experiments/LESSONS.md — durable cross-experiment findings