CLI Command
moro train
Executes local QLoRA fine-tuning on compiled datasets with hardware-aware memory limits and an integrated 5-level autonomous self-healing recovery engine.
Usage
moro train [OPTIONS]
Options & Flags
| Flag | Type | Default | Description |
|---|---|---|---|
| --config, -c | Path | ./moro.yaml | Path to recipe configuration file. |
| --dry-run | Boolean | false | Calculate memory tensors and print training plan without executing. |
| --auto-heal / --no-auto-heal | Boolean | true | Enable 5-level progressive OOM recovery and loss divergence rollback. |
| --resume | String | - | Resume from specified checkpoint ID or run ID. |
| --learning-rate, --lr | Float | - | Override learning rate defined in moro.yaml. |
| --batch-size | Int | - | Override per-device micro-batch size. |
Dry-Run Preview Example
$ moro train --dry-run Training Plan (Dry Run) ───────────────────────────────────────────── Model: Qwen/Qwen2.5-1.5B Adapter: LoRA (r=16, alpha=32, dropout=0.05) Effective Batch: 16 (micro_batch=2, accum=8) Learning Rate: 2e-4 (cosine scheduler) Estimated VRAM: 7.85 GB / 12.0 GB (36% headroom) Est. Duration: 18 minutes (1,200 steps) Zero OOM Status: VERIFIED SAFE