vae-opt

First audit every operator for cross-axis coupling: if no operator crosses the frame axis, shard along frames with no halo. If the decoder has recursive memory blocks, use state-carrying sliced decoding that passes each block's last-frame input activation across segment boundaries, then verify bitwise equivalence on CPU.

Run four verification gates: CPU whole-vs-sliced comparison requiring max|d| = 0, end-to-end md5 matching across requests, frame-level health metrics (high-frequency energy, frame diffs, frozen segments), and exchange-budget accounting. Note that md5 match alone only proves the pipeline output is unchanged, not decoder fidelity.

The choice is dictated by which operators couple which axes. Frame-axis sharding is preferred when nothing crosses time since it needs no halo and minimal communication; spatial tiling with halo sized to the receptive field works when only spatial convolutions couple pixels; recursive networks need state-carrying sequential slicing instead.

First prove sharding actually happened by printing state values (pieces count, per-shard shapes, whether state is None), since gating misreads cause silent fallback to full redundant decoding. Then run the CPU bitwise check; nonzero max|d| usually means state stored block output instead of input, missing clone, or cut points off latent-frame boundaries.

Skip it when decoding is under about 5 percent of end-to-end time, since DiT denoising and host overhead dominate. Also avoid it when the exchange budget fails: gathering fp32 video tensors costs gigabytes per card, so prefer uint8 exchange or writing tiles directly to the encoder.

No. Stock implementations with chunk_latents and trim values perform boundary artifact cropping, not independent computation. Shardability must be proven per operator by checking whether each output element depends on inputs outside the slice boundary.