The Agent Skills Directory

Resource Acquisition from Trusted Sources: The skill facilitates the download of machine learning models and optimized kernels from the Hugging Face Hub. These operations target well-known repositories and are essential for the performance optimization tasks the skill is designed for.
Dynamic Module Optimization: The provided scripts utilize dynamic patching of PyTorch modules at runtime. This allows for replacing standard operations with optimized Triton kernels without requiring permanent modifications to the installed libraries, which is a standard and legitimate practice for enhancing model performance on specific hardware like AMD GPUs.
Controlled Command Execution: The skill includes benchmarking scripts that execute local Python code for performance measurement. These scripts are transparent, use standard libraries, and follow the documented usage patterns without incorporating suspicious shell execution or privilege escalation vectors.
Secure Dependency Management: The skill defines its environment requirements in a standard requirements.txt file, referencing established machine learning and data processing packages. All identified packages are consistent with the skill's stated purpose of GPU-accelerated computing.

rocm-kernels