Serverless GPU (RunPod)

Custom RunPod serverless worker for FlixML. Bakes FLUX.2 models + your LoRAs into a Docker image.

What’s inside

Model	Path in image	Size
flux2_dev_fp8mixed.safetensors	models/diffusion_models/	~17 GB
mistral_3_small_flux2_bf16.safetensors	models/text_encoders/	~8 GB
flux2-vae.safetensors	models/vae/	~300 MB
Your LoRAs (docker/loras/*.safetensors)	models/loras/	varies

Build + Push

cd flixml/docker
IMAGE=<your-dockerhub-user>/flixml-flux2 ./build.sh

Requires:

• Docker Desktop or Docker Engine
• docker login to Docker Hub

Deploy on RunPod

1. Go to RunPod Console → Serverless
2. Create endpoint
3. Template → “Custom Image” → enter <your-dockerhub-user>/flixml-flux2:latest
4. GPU → pick your poison (RTX 4090 is cheapest, ~$0.50/hr)
5. Copy the Endpoint ID into FlixML .env:

   RUNPOD_API_KEY=your_key_here
   RUNPOD_ENDPOINT_ID=your_endpoint_id_here

Adding more LoRAs

Drop .safetensors files into docker/loras/ and rebuild:

cp new_lora.safetensors docker/loras/
cd docker && ./build.sh