Skip to main content

Ready-to-Run Templates

NovaPod Templates are pre-built, optimized container environments designed to bypass dependency installation, CUDA version matching, and kernel compilation issues. Every template launches in under 15 seconds with pre-warmed framework binaries and drivers.

Curated Official Templates​

1. PyTorch 2.4 + CUDA 12.4 Base​

  • Image: novapod/pytorch:2.4.1-cuda12.4-cudnn9
  • Default Ports: 8888 (JupyterLab), 22 (SSH)
  • Pre-installed: cuDNN 9, NCCL 2.20, FlashAttention-2, xFormers, Deepspeed, bitsandbytes, Accelerate, Wandb.
  • Best For: General deep learning, multi-GPU training, fine-tuning scripts, and research notebooks.

2. vLLM High-Throughput Serving Engine​

  • Image: vllm/vllm-openai:latest
  • Default Ports: 8000 (OpenAI REST API), 22 (SSH)
  • Key Features: PagedAttention, speculative decoding, continuous batching, and chunked prefill.
  • Quickstart: 
    # Inside your instance or configured via Startup Script
    vllm serve meta-llama/Meta-Llama-3-8B-Instruct --port 8000 --gpu-memory-utilization 0.95

3. ComfyUI + Flux.1 / SDXL Studio​

  • Image: novapod/comfyui:latest
  • Default Ports: 8188 (Web UI), 22 (SSH)
  • Key Features: Pre-configured nodes for Flux.1 Schnell & Dev, SDXL, ControlNet, Animatediff, and ComfyUI-Manager.
  • Access: Direct browser click opens the interactive graphical workflow canvas.

4. Ollama + OpenWebUI​

  • Image: ollama/ollama:latest
  • Default Ports: 11434 (API), 3000 (OpenWebUI)
  • Best For: Private, team-wide chat interfaces with zero terminal configuration required.

5. Unsloth & Axolotl LoRA Fine-Tuning​

  • Image: novapod/unsloth:latest
  • Default Ports: 8888 (JupyterLab), 22 (SSH)
  • Key Features: 2-5x faster backpropagation kernels with up to 80% lower VRAM overhead for Llama 3, Mistral, and Qwen models.

Custom Docker Images​

You can launch any public or private Docker image from Docker Hub, GitHub Container Registry (ghcr.io), or AWS ECR:

np launch \
  --gpu H100_SXM \
  --custom-image your-org/custom-cuda-runtime:v1.2 \
  --docker-args "--ipc=host --ulimit memlock=-1" \
  --ports "8000,8080"
Storage Retention

Files stored in the /workspace directory persist across instance Pause and Resume cycles. System root directories outside /workspace are ephemeral in Docker Pods. For full root persistence across the entire disk, use a Root KVM VM.