ai-toolkit

AI-Toolkit Workload

AI-Toolkit by ostris — a web-based interface for training LoRA adapters on diffusion models.

Deployment

The manifest at ai-toolkit-k8s.yaml creates:

• Namespace: ai-toolkit
• Deployment: Single replica running ostris/aitoolkit:latest
• Service: NodePort exposing the web UI on port 30675

GPU Assignment

The pod is pinned to the RTX 5090 (32 GB VRAM) via NVIDIA_VISIBLE_DEVICES UUID targeting. This is the only GPU with enough VRAM for training larger models. See GPU Assignment for the strategy.

Environment Variables

Variable	Value	Notes
AI_TOOLKIT_AUTH	YOUR_WEBUI_PASSWORD	Replace with your password
HF_TOKEN	YOUR_HF_TOKEN_HERE	Replace with your HuggingFace token
HF_HOME	/workspace/hf_cache	Cache directory for HF models
NVIDIA_VISIBLE_DEVICES	GPU-01b9a...	RTX 5090 UUID

Storage

Uses a hostPath volume at /data/ai-toolkit mounted to /workspace inside the container. This persists training configs, datasets, outputs, and cached models across pod restarts.

Resource Limits

resources:
  limits:
    nvidia.com/gpu: "1"
    memory: "30Gi"
  requests:
    nvidia.com/gpu: "1"
    memory: "8Gi"

The dshm (shared memory) volume is mounted at /dev/shm with a 1 GiB limit — required for PyTorch DataLoader workers.

Before Deploying

1. Replace YOUR_HF_TOKEN_HERE with your HuggingFace token
2. Replace YOUR_WEBUI_PASSWORD with a strong password
3. Ensure /data/ai-toolkit exists on the host
4. Verify the GPU UUID matches your target GPU (nvidia-smi -L)

kubectl apply -f ai-toolkit-k8s.yaml