|
AQIT 0.1.0
|
The binary is aqit. aquin is an alias with the same engine.
Python 3.9+. GPU stack is optional except for LLM paths.
aqit.cli.dispatch is a thin switch: AQIT verbs stay in this tree; everything else is forwarded to aquin.cli. Unknown verbs print Run aqit help.
| Command | Job |
|---|---|
| aqit help | Rich help (model-aware) |
| aqit status | Loaded model + local engine |
| aqit version | Installed version |
| aqit license | Apache 2.0 text + acceptance |
| Command | Job |
|---|---|
| aqit train <recipe.yaml> | DataRevision → train → Checkpoint → EvalGate |
| Command | Job |
|---|---|
| aqit load model <hf-id> | Load and keep resident in VRAM |
| aqit unload [--stop] | Free VRAM; --stop kills the daemon |
| aqit prompt <text> | Quick generation |
Do not background load. The model stays resident across later commands.
| Command | Job |
|---|---|
| aqit eval custom\|consistency\|suppress\|boundary | Probe suites |
| aqit check attention\|layer\|perturbation\|weight\|confidence\|trajectory | Health |
| aqit trace --prompt … --layer N | Causal / residual inspect |
| aqit diff weight\|sae\|residue --checkpoint <path> | Base vs checkpoint |
| aqit red-team | Adversarial probes |
| aqit benchmark --feature_idx N | Feature benchmarks |
Most GPU tools accept --check (JSON + PNG in the current directory) and --json (machine-readable stdout).
| Command | Job |
|---|---|
| aqit sae train\|align | Train or align sparse autoencoders |
| aqit info sae <id> | SAE metadata |
| aqit load sae <id> | Load catalog / user / path SAE |
| aqit feature locate\|logit\|neighbor | Feature tools |
| aqit activations capture\|replay | Activation store |
| aqit features compare | Compare captures |
| Command | Job |
|---|---|
| aqit steer --feature_idx N | Cheap intervention |
| aqit sweep | Strength grid |
| aqit multi-steer | Several features at once |
| aqit simulate | Forecast a train before GPU spend |
Renamed verbs print the new spelling and exit 1, e.g. feature-logits → aqit feature logit.