The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill instructions direct the agent to generate, compile, and execute custom Triton or TileLang kernels at runtime based on the model's architecture and performance hotspots. This involves the dynamic creation of executable code from generated source strings which are then compiled and run on the host system.
[INDIRECT_PROMPT_INJECTION]: The skill exposes a surface for indirect prompt injection by ingesting untrusted data from the user's codebase, including tensor shapes, strides, and code patterns, which directly influence the agent's code generation logic. The instructions do not mandate the use of boundary markers or sanitization for this external content.
Ingestion points: Project source code (via rg), model metadata (shapes, dtypes, strides), and performance trace logs.
Boundary markers: Absent; the skill relies on direct interpolation of extracted project details into kernel generation prompts.
Capability inventory: Execution of benchmarking scripts, runtime compilation of kernels (Triton/TileLang), and file system reads for code analysis.
Sanitization: Limited to numerical tolerance checks and verification against reference baselines after execution.
[COMMAND_EXECUTION]: The skill relies on various CLI tools and scripts to perform its functions. It uses rg (ripgrep) to search for synchronization points in the source code and utilizes vendor-specific tools like nvidia-smi and npu-smi to monitor hardware state. It also defines and executes benchmarking loops to measure model and kernel performance.

pytorch-perf-tuning