Autonomous GPU-kernel discovery & optimizer.

Technical Report

Ranked #1 on MLSys 2026 - FlashInfer AI Kernel Generation Contest for the DeepSeek Sparse Attention (DSA) track with an average speedup of 34.93x. Submissions can be found at:

Setup

Copy 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 format and can run without a local GPU on cloud using Modal. Requires Claude Code CLI.

# 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. To change the kernel you are implementing, please refer to the FlashInfer-Trace - Bring Your Own Kernel guide.

Launch the loop

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.

Architecture

For more details on the agentic loop, please refer to the technical report.

Agents:

• Profiler
• Research
• Workload inspector

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