Podstack Container Images

Podstack ships a catalog of pre-built, GPU-ready container images covering the most common AI/ML workloads. Pick one from the Template Catalog when creating a pod, or pull the image directly from Docker Hub.

All images are published at docker.io/manvarharsh/<image>:cuda12 (TensorFlow also has a cuda11 tag for older GPUs).

Conventions

Default Mount Path: /data

Every Podstack image expects your persistent storage at /data. Mount an NFS volume or attach a persistent disk here and your work survives pod restarts and recreations.

# Typical layout inside the pod
/data/
├── datasets/      # input data
├── models/        # checkpoints, weights
└── output/        # results

ComfyUI exception. For the ComfyUI image, do not change /data. Custom nodes, models, and workflows are stored under /data and need to live there to survive a restart. Change it and you’ll lose your custom-node setup every time the pod restarts.

CUDA Base

All GPU images are built on nvidia/cuda:12.4.1-cudnn-devel-ubuntu22.04 unless noted. They run on any modern NVIDIA GPU — RTX 30/40 series, L4, L40S, A100, H100. The cuda11 TensorFlow variant covers older Ampere/Volta/Turing hardware.

SSH + Service Architecture

Every image follows the same pattern:

• Port 22: SSH (controlled by ENABLE_SSH=true)
• Main service port: the application — JupyterHub (8000), Gradio (7860), the app’s native port (e.g. 8188 for ComfyUI, 11434 for Ollama)
• Entrypoint: /opt/podstack/scripts/entrypoint.sh wires env vars to services

Common Environment Variables

Image-specific variables (e.g. ENABLE_COMFYUI, COMFYUI_EXTRA_ARGS, VNC_PORT) are listed on each image’s page.

JupyterHub Auth

For images with JupyterHub, log in with:

• Username: any name (it’s just a session label)
• Password: your Podstack API token (psk_...) — generate one at Account > API Tokens

The token is validated against the Podstack backend.

Accessing Output Files in the Browser

Many workflows produce output files (renders, generated images, processed video) under /data/output/. To browse those files in your browser:

1. Expose a Port on the Pod

1. Open the pod’s detail page in the portal.
2. Click Edit Deployment.
3. Add a new exposed port (e.g. 8080).
4. Save. A new URL mapped to that port appears on the pod page.

2. SSH into the Pod and cd to the Output Directory

ssh root@<pod-subdomain>.cloud.podstack.ai
cd /data/output

Use whichever directory your app writes to — /data/output, /data/comfyui/output, /data/renders, etc.

3. Start Python’s Built-In HTTP Server

python3 -m http.server 8080

You should see:

Serving HTTP on 0.0.0.0 port 8080 (http://0.0.0.0:8080/) ...

4. Open the Mapped URL

Open the URL surfaced under the pod’s exposed port (from step 1) — it routes traffic to port 8080 inside the container. You’ll see a directory listing with every file in /data/output/ available to download or preview.

Press Ctrl+C in the SSH session to stop the server when you’re done.

Note: This serves your files publicly to anyone with the URL. Don’t expose sensitive data this way, or stop the server immediately after you’re done.

Image Catalog

Image Generation

• a1111 — Automatic1111 Stable Diffusion WebUI
• sd-forge — SD WebUI Forge (faster A1111 fork)
• comfyui — ComfyUI node-graph image/video generator
• kohya-ss — Kohya LoRA / fine-tune trainer

LLM Training

• axolotl — Axolotl LLM fine-tuning framework
• llamafactory — LLaMA-Factory GUI training
• unsloth — Unsloth 2–5× faster fine-tuning
• unsloth-studio — Unsloth Studio (managed UI)

LLM Serving & Inference

• vllm — vLLM high-throughput inference server
• sglang — SGLang structured-generation server
• ollama — Ollama local LLM runner
• text-gen-webui — oobabooga text-generation-webui
• tei — Hugging Face Text Embeddings Inference
• triton — NVIDIA Triton Inference Server

ML Frameworks & Notebooks

• pytorch — PyTorch + Jupyter
• tensorflow — TensorFlow + Jupyter (CUDA 11 and 12)
• jupyterlab-gpu — Bare JupyterLab on GPU
• rapids — NVIDIA RAPIDS (cuDF, cuML)
• tensorrt — NVIDIA TensorRT optimizer / runtime

Speech & Audio

• cloudtts — MOSS-TTS + Chatterbox TTS
• whisperx — WhisperX speech-to-text + diarization

Science & HPC

• alphafold — AlphaFold protein structure prediction
• gromacs — GROMACS molecular dynamics
• octave — GNU Octave numerical computing

Video, Graphics & Vision

• cloudblenderrender — Blender 4.4 GPU rendering
• ffmpeg-gpu — FFmpeg with NVENC/NVDEC
• sam3 — Meta Segment Anything Model 3

Base

• ubuntu-ssh-cuda — Bare Ubuntu + CUDA + SSH (BYO stack)