diffusers

29bd6e06 - Add slow modular-pipeline integration tests with declared offload behavior

Add slow modular-pipeline integration tests with declared offload behavior Every recorded offload event now carries the device's cumulative peak-memory reading (never reset by the offloader, so deltas between consecutive events attribute activation peaks to the interval between moves). New ModularPipelineIntegrationTesterMixin runs a modular pipeline against its real checkpoint: subclasses declare the expected offloading behavior per simulated card (offload counts, OOM-retry counts, final device per model) and test_auto_cpu_offload_on_cards verifies the run matches the spec and leaves the output untouched. The docstring carries the discovery recipe for transcribing a new pipeline's behavior from the offload record. TestZImageModularIntegration exercises Z-Image-Turbo on 32/24/16/10GB cards (no offloading / text encoder yields / transformer also yields / run-alone escape hatch) plus text-to-image and image-to-image slice checks. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>