Summary

• Scope the diffusion-stage transformer cache to its single intended reuse: carrying an identical transformer from the first diffusion stage into the second.
• Evict the cached transformer as soon as the reuse-hit context exits, before VAE/audio decoder construction.
• Keep generation-start eviction as a defensive backstop for cancellation, failure, or non-identical follow-up stages.
• Enable the corrected cache by default and regenerate the OpenAPI artifacts.
• Add regression coverage proving a consumed hit is released and a third identical stage rebuilds normally.

Root cause

The generation-scoped cache retained the reused transformer after the second diffusion stage. On Apple Silicon this allowed roughly 37 GiB of transformer state to overlap tiled VAE decoding. A live 1280x704 run with that lifetime peaked at 46.7 GiB of MPS driver memory and still had 35.4 GiB torch-allocated after output encoding. The cache only needs to survive the boundary between the two identical diffusion stages.

Apple Silicon HD A/B

Workload: Fast pipeline, 1280x704, 121 frames, 24 fps (5.04 seconds), fixed prompt and seed. Two fresh runs per condition.

All four measured MP4s had the same SHA-256 (7e82...78aa); the comparison reported 99 dB PSNR. The timing result is directional because each condition has only two runs; the memory/lifetime fix is the primary motivation.

Validation

• pnpm backend:test — 568 passed
• pnpm typecheck — TypeScript clean; Pyright 0 errors/0 warnings
• pnpm build:frontend — production build passed
• pnpm openapi:check — generated schema/types current

Review note

This is a draft so the newly enabled default can receive cross-platform/CUDA validation before merge.