Sartiq AI Toolkit

Fork of ostris/ai-toolkit with custom modifications for Shootify's training pipeline.

Deployed at: https://ai-toolkit-01.shootify.io/dashboard

---

Custom Modifications

1. Resolution Bucketing (4:5 Portrait)

Replaced the original resolution system with explicit [width, height] bucket pairs and added 4:5 portrait aspect ratio buckets optimized for garment/fashion training:

• 8 new 4:5 buckets: 960x1200 through 1792x2240
• New ResolutionSelector UI component with per-bucket width/height editing
• TypeScript bucket utility mirroring Python bucket generation logic
• Migration support for old config formats (single int, flat array)

Files changed:

• toolkit/buckets.py — new 4:5 portrait buckets
• toolkit/config_modules.py — resolution format as [w, h][]
• toolkit/dataloader_mixins.py — support exact [w, h] bucket dimensions
• ui/src/components/ResolutionSelector.tsx — new component
• ui/src/utils/buckets.ts — TypeScript bucket utility
• ui/src/app/jobs/new/SimpleJob.tsx — use ResolutionSelector
• ui/src/app/jobs/new/jobConfig.ts — resolution migration
• ui/src/app/jobs/new/page.tsx — simplified view
• ui/src/types.ts — updated resolution type

2. HuggingFace Download Fix

hf_transfer (Rust-based download accelerator) has a bug that stalls downloads at the 2^34 byte boundary (~16GB). Disabled it and Xet storage protocol to use plain HTTP instead.

Environment variables added to job spawn:

• HF_HUB_ENABLE_HF_TRANSFER=0
• HF_HUB_DISABLE_XET=1
• HF_HUB_DOWNLOAD_TIMEOUT=300
• HF_HOME=/data/.cache/huggingface

File changed:

• ui/cron/actions/startJob.ts

3. Dockerfile

Clones from shootify-io/ai-toolkit fork instead of upstream ostris repo.

File changed:

• docker/Dockerfile

4. Dev Server Config

• Added allowedDevOrigins for ai-toolkit-01.shootify.io
• Set dev port to 8675
• Removed turbopack

Files changed:

• ui/next.config.ts
• ui/package.json

---

Deployment

Infrastructure

• Host: GCP VM (A100-SXM4-80GB, CUDA 12.4)
• URL: https://ai-toolkit-01.shootify.io/dashboard
• Working directory: /home/info/shootify-deployment/ai-toolkit/

Running (Dev Server, no Docker)

The app runs directly on the host via npm run dev (Next.js + cron worker via concurrently) on port 8675.

Start:

cd /home/info/shootify-deployment/ai-toolkit/ui
nohup npm run dev > /tmp/dev-server.log 2>&1 &

Stop:

pkill -f "next dev --port 8675"
pkill -f "ts-node-dev.*worker"
pkill -f "concurrently.*WORKER"

Logs: /tmp/dev-server.log

There is also a systemd service at /etc/systemd/system/ai-toolkit.service.

Python Environment

.venv is a conda environment with Python 3.12 created at the repo root:

conda create -p /home/info/shootify-deployment/ai-toolkit/.venv python=3.12 -y
.venv/bin/pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
.venv/bin/pip install -r requirements.txt

The startJob.ts worker detects .venv/bin/python automatically.

Data Paths

Path	Contents
/home/info/shootify-deployment/ai-toolkit/output/	Training outputs (logs, checkpoints, samples)
/home/info/shootify-deployment/aitk_db.db	SQLite database (jobs, settings)
/data/.cache/huggingface/	HuggingFace model cache
/home/info/shootify-deployment/ai-toolkit/datasets/	Training datasets

---

Upstream

For general AI Toolkit documentation, features, and supported models, see the original repo: ostris/ai-toolkit