1-Click Launch Recipes for Vast.ai

Docker & SSH

Copy-paste deployment recipes for ComfyUI, vLLM, Ollama, and PyTorch

Live Rates Cost Simulator Launch on Vast.ai

Developer Launch Recipes

1-Click Docker & SSH Recipes for Vast.ai

Tested terminal commands for immediate deployment on NVIDIA GeForce RTX 3090 instances.

Container Ready

# 1. SSH into your Vast.ai instance:
# ssh root@<instance_ip> -p <ssh_port>

# 2. Launch ComfyUI with FLUX.1 & SDXL model cache:
docker run -d --gpus all --name comfyui \
  --restart unless-stopped \
  -p 8188:8188 \
  -v /workspace/ComfyUI/models:/root/ComfyUI/models \
  -v /workspace/ComfyUI/output:/root/ComfyUI/output \
  yanwk/comfyui-boot:cu124

# Access ComfyUI in your browser at: http://<instance_ip>:8188

💡 Tip: Make sure your security group / host port mapping allows ingress on the selected port.Open Vast.ai Console

1. ComfyUI + FLUX.1 Persistent Cache

Port 8188

Always mount your models directory to a persistent volume (e.g., /workspace/ComfyUI/models). This ensures that heavy 24GB checkpoint weights (FLUX.1 [dev], SDXL, SD 1.5) remain saved when you stop your instance, preventing hours of repetitive downloads.

# Mount models volume into container:
-v /workspace/ComfyUI/models:/root/ComfyUI/models \
-v /workspace/ComfyUI/output:/root/ComfyUI/output

2. High-Throughput vLLM Server

Port 8000

Deploy Meta-Llama-3.3-70B or DeepSeek-R1 with OpenAI-compatible API drop-in replacement. Set --tensor-parallel-size equal to your allocated GPU count (e.g., 2 or 4 GPUs for 70B models).

vllm/vllm-openai:latest \
  --model meta-llama/Llama-3.3-70B-Instruct \
  --tensor-parallel-size 2 \
  --gpu-memory-utilization 0.95

3. Port Forwarding & Networking

SSH Tunneling

If Vast.ai uses NAT without direct public IP mapping, use an SSH local port forward to access web UIs securely on your local browser:

# Forward remote ComfyUI (8188) to local localhost:8188:
ssh -L 8188:localhost:8188 root@<host_ip> -p <ssh_port>

4. Zero-Egress Cloud Backup (R2 / S3)

rclone

Avoid 24/7 idle stopped storage fees on Vast.ai by syncing your fine-tuned weights and model checkpoints directly to Cloudflare R2:

# Fast multi-threaded sync with rclone:
rclone copy /workspace/output r2:my-bucket/checkpoints --transfers 8