Hermes RunPod Pipeline

Compute infrastructure orchestrator for AI image generation on the RunPod GPU cloud. Automated pod lifecycle management with multi-region GPU selection, model deployment, and inference workflow automation.

Overview

This toolchain manages the full lifecycle of GPU-accelerated image generation on RunPod:

• Pod orchestration — Spin up GPU pods across 3 regions (Americas, Europe, Asia) with automatic failover. Supports RTX 4090, RTX 5090, and A100 80GB on Secure Cloud.
• Model deployment — Download and configure diffusion models from HuggingFace at pod boot time. Parallel download architecture achieves <30s cold start for full model stacks.
• ComfyUI automation — Deploy ComfyUI, submit workflow JSONs, poll for completion, and retrieve results — all through SSH-tunnelled API calls.
• LoRA training — Fine-tune diffusion models on custom datasets using AI-Toolkit, with automatic dataset preparation and weight distribution.

Quick Start

# Install dependencies
pip install -r requirements.txt

# Configure credentials (see docs/credentials.md)
cp .env.example .env
# Edit .env with your RunPod API key and HuggingFace token

# Generate images with a trained LoRA
python generation/comfyui_gen.py \
  --lora "https://files.catbox.moe/xxxxx.safetensors" \
  --lora-strength 0.85 \
  --prompts "subject, photorealistic, detailed" \
  --count 5

Features

• Multi-region GPU selection — Automatically finds available GPUs across Americas → Europe → Asia with 3×3 fallback (3 GPU types × 3 regions)
• Parallel model loading — Downloads FLUX transformer, text encoders, VAE, and LoRAs concurrently while ComfyUI installs its dependencies
• MD5 integrity verification — All model weights and output images are checksum-verified before use
• Cost-optimized — Pods terminate automatically after generation completes. Typical sessions cost $0.08–0.10.
• Composable — Infrastructure modules (GPU selection, SSH, RunPod API) are reusable across training and generation pipelines

GPU Selection Logic

GPU	Cloud	Hourly	Region Priority
RTX 4090 24GB	Secure	~$0.69/hr	Americas → Europe → Asia
RTX 5090 32GB	Secure	~$0.69/hr	Americas → Europe → Asia
A100 80GB	Secure	~$1.19/hr	Americas → Europe → Asia

Fallback order: 4090 → 5090 → A100, each tried across all 3 regions.

Project Structure

├── generation/
│   ├── comfyui_gen.py          # Image generation pipeline
│   ├── setup_models.sh         # Parallel model download script
│   └── workflow_builder.py     # ComfyUI workflow generator
├── training/
│   ├── train_lora.py           # LoRA fine-tuning pipeline
│   ├── setup_training.sh       # Training environment setup
│   └── config_templates/       # AI-Toolkit YAML templates
├── infrastructure/
│   ├── gpu_selector.py         # GPU + region selection
│   ├── runpod_utils.py         # RunPod API wrappers
│   └── ssh_utils.py            # SSH connection management
├── registry/
│   └── lora_manifest.json      # Trained model registry
└── docs/
    ├── credentials.md          # API key setup guide
    └── troubleshooting.md      # Known issues and fixes

Requirements

• Python 3.10+
• RunPod account with funded balance
• HuggingFace account with read token
• SSH key registered with RunPod

Configuration

See docs/credentials.md for setting up API keys and authentication.

Cold Start Performance

Phase	Time
Pod provisioning + GPU allocation	30–60s
Model downloads (parallel, 22.5 GB)	~30s
ComfyUI boot + VRAM load	5–10s
First inference (28 steps, 832×1216)	10–15s
Total	~1.5–2 min

License

MIT