memory-safety findings in llm.c CUDA kernels

Hello maintainers,

I am working on a research project for automating fuzz testing of CUDA libraries using cuFuzz (https://github.com/NVlabs/cuFuzz), and I came across llm.c while evaluating open-source CUDA projects.

While experimenting with llm.c, we found some memory-safety issues affecting CUDA kernels and related host-side launch paths. At a high level, these include out-of-bounds accesses, integer-overflow-driven unsafe launches, null-pointer/invalid-pointer cases, and synchronization/race issues.

We have prepared a small report package with per-finding summaries, reproducers, triggering inputs where applicable, sanitizer output, CWE labels, and suggested fix directions. Would you prefer that we attach the reports here, or is there a different communication channel you would like us to use?

Thank you,
Mohamed