---

name: vit-pytorch description: "Routes vit-pytorch Vision Transformer model construction, variable-resolution and video inputs, pretraining wrappers, attention introspection, and package troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: MIT

vit-pytorch

Use this repo skill when a task names vit-pytorch, vit_pytorch, lucidrains' Vision Transformer package, or a model/workflow implemented by that package: ViT, SimpleViT, NaViT, ViViT, CCT, CrossViT, MAE, Dino, Recorder, Extractor, and related image/video transformer variants.

Quick setup and import check

Install the public package in the user's Python environment:

pip install vit-pytorch

Then run a minimal import/forward check:

import torch
from vit_pytorch import ViT

model = ViT(image_size=32, patch_size=8, num_classes=7,
            dim=32, depth=1, heads=2, dim_head=16, mlp_dim=64)
logits = model(torch.randn(2, 3, 32, 32))
assert logits.shape == (2, 7)

For a broader package health check, run the bundled helper after installing the package:

python scripts/check_vit_pytorch_install.py --run-smoke

The package exposes Python modules only; no console CLI entry points were found in this snapshot.

Route by user intent

• Fixed-size 2D image classifier or architecture choice: read sub-skills/image-architectures/SKILL.md. Use it for plain ViT / SimpleViT, CCT, CrossViT, PiT, LeViT, CvT, MaxViT, XCiT, JetViT, ViT-5, and image-only variants such as patch dropout or patch merger.
• Variable-resolution images, NaViT grouping, 1D/3D/N-D tensors, or video: read sub-skills/variable-shapes-video/SKILL.md. Use it for NaViT, nested tensor variants, vit_3d, simple_vit_3d, cct_3d, ViViT, 1D models, N-D models, and video wrappers.
• Loss wrappers, pretraining, distillation, adaptation, or fine-tuning: read sub-skills/pretraining-and-adaptation/SKILL.md. Use it for DistillWrapper, MAE, SimMIM, MPP, MP3, DINO, EsViT, Learnable Memory ViT, LeJEPA, VAT, VAAT, WWT, and decorrelation auxiliary loss.
• Attention maps, embeddings, hooks, efficient/custom transformer injection, or performance wrappers: read sub-skills/introspection-and-customization/SKILL.md. Use it for Recorder, Extractor, efficient.ViT, parallel_vit, and simple_flash_attn_vit.

Operating rules

1. Identify the user's tensor layout before choosing a route: image (batch, channels, height, width), NaViT image lists (channels, height, width), video (batch, channels, frames, height, width), 1D (batch, channels, seq_len), or N-D (batch, channels, *input_shape).
2. Start with tiny random tensors and small constructor dimensions before copying README-scale examples. Most helper scripts use dim=32, depth=1, and CPU-safe tensors to validate shape and wrapper wiring.
3. Treat optional dependencies as workflow-specific. torch, torchvision, and einops are base dependencies; torchaudio is needed for VAAT, while Accelerate, W&B, Kaggle data, external transformer packages, or checkpoints belong only to explicit full-training or research-idea requests.
4. Do not run dataset downloads, notebooks, long training loops, or checkpoint loaders as default smoke checks. Use the bundled helpers first, then ask for data/runtime approval if the user requests full training.
5. If a workflow fails, read references/troubleshooting.md and the nearest sub-skill troubleshooting reference before changing packages or model families.
6. If working against a live checkout rather than an installed release, read references/repo-provenance.md to decide whether this skill should be refreshed.

Shared references and scripts

• references/api-reference.md — top-level package facts, important exports, verified signatures, and optional dependency notes.
• references/troubleshooting.md — cross-cutting install/import, backend, optional dependency, and version-fragile wrapper guidance.
• references/repo-provenance.md — source snapshot and refresh baseline.
• references/repo-routing-metadata.json — structured routing metadata for managed repo-skill import.
• scripts/check_vit_pytorch_install.py — package import, metadata, optional module, and minimal forward smoke checks.