- Encode prompts under no_grad so cached embeddings hold no autograd graph that would keep the released text encoder alive. - Release and reload the pipeline when a prompt stage fails after the DiT was dropped, instead of failing every later request on transformer None. - Hash the bundle's pipeline_moss_image.py that the MPS path executes, and record mps_runtime.py separately, instead of hashing only the wrapper. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>