---

name: minivit description: "Routes MiniViT Mini-DeiT and Mini-Swin compression, distillation, command-construction, dataset-layout, and optional RPE/custom-op workflows." disable-model-invocation: true metadata: disco-role: operating license: MIT

MiniViT

Use this sub-skill when the user asks about MiniViT workflows for Mini-DeiT or Mini-Swin: model/variant selection, ImageNet layout validation, MiniViT weight multiplexing, distillation command construction, evaluation command construction, 224-to-384 Mini-Swin finetuning, or optional RPE/custom-op status.

Route first

User task	Read/run
Build a Mini-DeiT or Mini-Swin command template	scripts/build_minivit_command.py, then references/command-reference.md
Choose a model variant, YAML config, or checkpoint family	references/api-reference.md and references/command-reference.md
Validate ImageNet folders or tar archives	references/workflows.md and root scripts/check_dataset_layout.py
Diagnose teacher checkpoints, rpe_ops, Mini-Swin configs, Apex, or layer-list mismatches	references/troubleshooting.md

Safe operating pattern

1. Confirm whether the user means Mini-DeiT or Mini-Swin. Generic iRPE integration belongs to the irpe sub-skill; TinyViT and EfficientViT belong to their own sub-skills.

2. Validate the ImageNet root before constructing expensive launchers:

   python ../../scripts/check_dataset_layout.py --root /path/to/ImageNet --kind imagenet1k
   

3. Generate, review, and edit a command template instead of improvising distributed flags:

   python scripts/build_minivit_command.py --workflow mini-swin-train --variant tiny --data-path /path/to/ImageNet --output /path/to/output --teacher /path/to/teacher.pth --amp-opt-level O0
   

4. Treat printed commands as templates. The helper never starts distributed training, downloads data, or writes checkpoints; commands printed by the helper may do those things if the user later runs them.

Native safe checks

• Mini-DeiT parser help is safe in an environment with timm installed.
• mini_deit_tiny_patch16_224 CPU instantiation is safe and was verified in the inspection environment.
• Mini-Swin config parsing with an explicit --cfg is safe; full Mini-Swin training/evaluation requires CUDA, DistributedDataParallel, and NCCL setup.

Boundaries

• Covered here: Mini-DeiT and Mini-Swin compression/distillation/evaluation command construction and troubleshooting.
• Not covered here: generic iRPE API integration, TinyViT sparse-logit workflows, EfficientViT classification/downstream workflows, or Cream/AutoFormer NAS search.
• Keep runtime guidance self-contained: use bundled references/helpers rather than source-repo documentation links.