Model2Kernel: Model-Aware Symbolic Execution For Safe CUDA Kernels

Researchers developed Model2Kernel, a system that automatically detects memory safety bugs in CUDA kernels used for large language model inference. The system discovered 353 previously unknown bugs across popular platforms like vLLM and Hugging Face with only nine false positives.

• →Model2Kernel is the first practical system for automatically verifying memory safety of CUDA kernels in LLM inference systems.
• →The system uses model-aware dynamic analysis and CUDA-specialized symbolic execution to detect memory bugs.
• →Testing on popular platforms revealed 353 previously unknown bugs with high accuracy (only 9 false positives).
• →Memory safety bugs in GPU kernels can corrupt model weights, crash services, or enable adversarial attacks.
• →Existing techniques failed to handle variable-length kernel inputs or had prohibitive overhead for production use.