The Agent Skills Directory

Functional Code Execution: The benchmarking and profiling scripts utilize dynamic loading to execute generated Triton kernels. This is a core part of the optimization loop, allowing for the verification of performance and numerical correctness of new implementations during the development process.
Subprocess Integration: The skill interfaces with external profiling utilities, such as Intel VTune, using subprocess.run. These operations are used to collect hardware performance counters and are appropriate for the skill's role as an optimization framework.
External Resource Usage: The skill utilizes the ai-bench framework, which is referenced from a reputable external repository specialized in performance benchmarking for hardware accelerators.
Module Patching Pattern: Included examples demonstrate how to integrate optimized kernels into existing models by replacing specific operations at runtime. This practice is common in performance engineering to allow for implementation swaps without modifying library source code.
Static Analysis Note: While some automated tools may flag the use of .eval() in the scripts, this refers to the standard PyTorch nn.Module.eval() method used to set models to evaluation mode for inference, rather than the built-in Python eval() function.

xpu-kernels