TRT Perf Analysis

Workflow

Use <runner> below as scripts/run.sh on Unix-like systems or scripts\run.cmd on Windows. Set SKILL_PYTHON to override Python 3.8+ discovery.

Treat each step as a gate. On failure, stop, report the error, clean up temporary outputs, and do not create or preview a report.

1. Resolve inputs in this order:

   1. Use JSON pairs explicitly provided by the user; preserve their pairing. Filenames may be arbitrary.
   2. For a provided folder, find complete pairs there. Match conventional layers_<label>.json and profile_<label>.json names first; use JSON structure and layer names for nonstandard names. Handle component subfolders separately.
   3. With no explicit path, infer only from clear context: attachments, discussed paths, or the active workspace. Fail if no unambiguous pair is found.

   Require at least one readable layer-info and profile/latency pair. Accept raw detailed TensorRT inspector output as either a layer array or a TensorRT 11 object containing Layers. Do not use analyze-data.json as layer-info input. Reject missing, ambiguous, layer-only, or profile-only inputs.

2. Analyze in staging. Run scripts/analyze_trt_perf.py before interpretation. Use repeated --data <layer.json> <profile.json> for explicit or nonstandard pairs; use folder mode only for conventionally named pairs. Pass a user-provided model name with --model-name; otherwise use analyzer inference. Write analyze-data.json outside the final report folder and use it as the source for all conclusions.