sabertooth-123/llm_instrumentation_tracing_replay

Transformer Inspector loads Hugging Face causal language models, repeatedly runs inference, and captures layer-level activity for interactive inspection. The main path is CLI configuration → model loading and hook registration → inference and analysis → bounded snapshot storage → terminal or browser views. Analysis includes activation metrics and anomalies, attention, residual-stream and token-prediction features, and optional head ablation; the web mode also supports comparing two models. The README documents these modes, but the web page source and some runtime code were not sampled, so uncertain frontend and inference-call details are omitted.