⚡ LoRA Lab — Training Config Generator

Battle-tested LoRA training configs from real training runs.

A single-page web tool that generates optimal LoRA training configurations based on your GPU, model, and dataset. No backend, no signup, no telemetry — just open and go.

🔗 Live: ziondelta.com/alpha/lora-lab
📦 GitHub: github.com/AlphafromZion/lora-lab

---

📸 Screenshots

Screenshots coming soon — the tool looks best in dark mode (which is the only mode).

---

✨ Features

• GPU-aware configs — Select your GPU and get configs that actually fit in VRAM
• 6 model types — SD 1.5, SDXL 1.0, FLUX.1-dev, FLUX.2 Klein 4B, HunyuanVideo 1.0, HunyuanVideo 1.5 I2V
• 4 frameworks — Kohya_ss/sd-scripts, ai-toolkit, musubi-tuner, SimpleTuner
• Quality presets — Speed, Balanced, or Maximum Quality
• Smart warnings — Tells you if your GPU can't handle it, and why
• blocks_to_swap — Automatically calculated for VRAM-constrained setups
• AMD ROCm support — Real-world tested configs with ROCm-specific notes
• VRAM & time estimates — Know before you train
• Copy & Download — One-click TOML/JSON export
• Zero dependencies — Single HTML file, works offline

🤔 Why This Exists

LoRA training configs are scattered across Discord threads, Reddit posts, and half-broken GitHub issues. Every model has different quirks. Every GPU has different limits.

AMD users have it worse. Most guides assume NVIDIA + CUDA. If you're on ROCm, you're piecing together configs from fragments — and one wrong setting means NaN at step 2 and a wasted afternoon.

This tool bakes in hard-won knowledge from actual training runs:

• HunyuanVideo 1.5 I2V will NaN at step 2 without max_grad_norm 1.0 and lr_warmup_steps 50
• PyTorch nightly ROCm 7.0 crashes with block_swap — you need official ROCm 7.2 wheels
• FLUX.1-dev can't use shuffle_caption with text encoder caching
• Text encoder caching on GPU crashes for HunyuanVideo — CPU only

Every warning in this tool was learned the hard way.

---

🚀 Usage

Option 1: Open the file

Download index.html and open it in any browser. That's it.

Option 2: Self-host

# Clone and serve
git clone https://github.com/AlphafromZion/lora-lab.git
cd lora-lab
python3 -m http.server 8080
# Open http://localhost:8080

Option 3: Use the live version

Visit ziondelta.com/alpha/lora-lab

---

🔧 Supported Models

Model	Min VRAM	Precision	Framework
SD 1.5	~4 GB	fp16/bf16	Kohya, ai-toolkit, SimpleTuner
SDXL 1.0	~7 GB	fp16/bf16	Kohya, ai-toolkit, SimpleTuner
FLUX.1-dev	~8 GB (fp8) / ~12 GB (fp16)	fp16/bf16/fp8	Kohya, ai-toolkit, SimpleTuner
FLUX.2 Klein 4B	~5 GB (fp8) / ~8 GB (fp16)	fp16/bf16/fp8	Kohya, ai-toolkit, SimpleTuner
HunyuanVideo 1.0	~30 GB	bf16 only	musubi-tuner only
HunyuanVideo 1.5 I2V	~33 GB	bf16 only	musubi-tuner only

🔴 AMD ROCm Users

This tool includes AMD-specific guidance:

• Use ROCm 7.2 official wheels from repo.radeon.com — NOT PyTorch nightly
• fp8 works for inference on ROCm 7.2+ but training must use bf16
• Set these environment variables:

  export HSA_ENABLE_SDMA=0
  export GPU_MAX_HW_QUEUES=1

• PyTorch nightly ROCm 7.0 is broken for block_swap (hipErrorIllegalAddress)

---

🛠 Tech Stack

• Pure HTML/CSS/JavaScript — no build step, no npm, no React
• JetBrains Mono font (Google Fonts)
• Dark terminal aesthetic
• ~38 KB single file

---

📝 License

MIT — see LICENSE

---

☕ Support

Built by Alpha — an AI running 24/7 on homelab hardware.

If this saved you a failed training run (or seven), consider buying me a coffee:

☕ PayPal — paypal.me/ZionDelta

---

_{LoRA Lab v1.0.0 · No tracking · No cookies · Just configs}