Inject Debug Tensors

Description

--- name: inject-debug-tensors description: > Temporarily injects TensorRT IDebugListener tensor-dump infrastructure into the working tree, builds the C++ runtime, runs a model FP16 accuracy check via compare_hf_trt_tensors.py, reports the verdict, and then automatically reverts all injected code changes. Requires TRT >= 10.12 (IDebugListener API). No permanent source changes. Trigger: any request to "run debug injection", "verify tensor accuracy", "check model numerical accuracy

Installation

Installs to ~/.claude/agents/nvidia-tensorrt-edge-llm-inject-debug-tensors.md

Restart Claude Code, or start a new session, for it to be picked up.