Home Papers Patents Talks Posts Experience Projects Contact
All publications

GPU Kernels

GPU Kernel Correctness

The Correctness Illusion, op-schema-aware seeded fuzzing, and the 26-op gpuemu corpus for LLM-generated GPU kernels.

FAQ

What is the Correctness Illusion?

The Correctness Illusion: LLM-generated GPU kernels pass the standard correctness test (KernelBench) and are still wrong in production. Published June 2026 on arXiv (2606.20128). The standard test under-specifies the input distribution, so kernels that pass can still fail on real workloads.

What is op-schema-aware seeded fuzzing?

Op-schema-aware seeded fuzzing is our technique for finding bugs in LLM-generated GPU kernels. We take the op schema (input shape, dtype, constraints) and generate seeds that exercise boundaries. 4-8x more bugs in less time than uniform sampling. arXiv:2606.27396.

Where can I get the 26-op corpus?

On Hugging Face at huggingface.co/dipankarsarkar/gpuemu-corpus. The harness is on GitHub at github.com/sarkar-dipankar/gpuemu-corpus. The Hugging Face Space at dipankarsarkar (The Correctness Illusion) lets you run it interactively.

Related

  • dipankar.org — strategic studies + advisor practice (broad public persona)
  • dipankar.name — engineering work + AI agent infrastructure (canonical hub)
  • dipankar.co — fractional CTO + consulting practice