Debugger
Debugger mode is a tiled engineer dashboard for deep model inspection. It is organized into panels that can be rearranged. All data comes from real model calls to /debug/* and /ablate/analyze.
Panels
| Panel | What it does |
|---|---|
| Breakpoints | Set layer/head breakpoints. Playback pauses when the specified component executes. |
| Flame graph | Per-op timing breakdown showing host-side execution times. PROXY label on MPS/CUDA (no device sync). |
| Anomaly sentinels | Automatic alerts for NaN, Inf, near-zero activations, and attention collapse patterns. |
| Watch expressions | Monitor specific tensor values across layers or steps. |
| Layer table | Per-layer summary: residual delta, norm magnitude, top attention weight. |
| Head grid | All 14 × all layers attention heatmaps in a compact grid view. |
| Ablation | Zero out a selected head or layer and compare before/after logit distributions. Uses POST /ablate/analyze. |
| Induction-head lab | Identifies candidate induction heads by their characteristic copy-suppression attention pattern. |
| Trace replay | Replay any saved trace. Branch from a replay to run a modified forward pass and compare the two traces side by side. |
| Console REPL | Send arbitrary commands to the debug API. Useful for extracting tensor values not exposed by the standard panels. |
Activation patching
Warning
The “Activation Patching” panel title is a known inaccuracy — the panel shows a logit-lens trajectory (a correlational measurement), not causal activation patching. Tracked in issue #75.