pipe = DiffusionPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16,
)
pipe.enable_model_cpu_offload()
# Instead of full compilation, we apply regional compilation
# here to take advantage of `fullgraph=True` and also to reduce
# compilation time. More details can be found here:
# https://hf.co/docs/diffusers/main/en/optimization/fp16#regional-compilation
pipe.transformer.compile_repeated_blocks(fullgraph=True)
image = pipe(**pipe_kwargs).images[0]