Fix: ensure_on_device before each inference to prevent text_encoder drift

Explicitly move all pipeline components to CUDA, use inference_mode

Remove generator entirely to fix CUDA/CPU device mismatch

Fix generator - remove device spec, use official example pattern

Fix CUDA device mismatch - resize image and add autocast

Update README.md

Use public pytorch image, install cosmos-guardrail with --no-deps

Add FastAPI handler

Add Dockerfile using NVIDIA cosmos container

initial commit