Hi @shreshthsaini, thanks for the PR! It does not appear to link an issue it fixes. If this PR addresses an existing issue, please add a closing keyword (e.g. Fixes #1234) to the PR description so the issue is linked. See the contribution guide for more details. If this PR intentionally does not fix a tracked issue, a maintainer can add the no-issue-needed label to silence this reminder.

Please note that PRs without a linked issue are likely to be automatically closed 10 days after this notice.

Once the PR links an issue (or gets the no-issue-needed label), you can ignore this message — it stays here as a comment, but it no longer applies.

Self-review notes and test log, per the contribution guide.

Test

CPU check with the tiny Wan components from tests/pipelines/wan/test_wan.py (12 steps, threshold=0.5 so the tiny model actually skips, a mean-pixel verifier, N=3):

video, seed, scores = cached_search(pipe, prompt, verifier, n=3, threshold=0.5, num_inference_steps=12, **kw)
ref = pipe(prompt=prompt, num_inference_steps=12, generator=torch.Generator().manual_seed(seed), **kw).frames[0]

Output:

last-block calls during search + commit: 30   (4 full runs would be 96)
committed video identical to plain run: True
cache hooks removed after search: True

Self-review notes

• The script uses only public Diffusers API (enable_cache(FirstBlockCacheConfig(...)), disable_cache) and relies on maybe_free_model_hooks resetting the cache state at the end of every pipeline call, so candidates do not share state.
• transformer_2 (Wan 2.2) is cached as well when present.
• The FirstBlockCache threshold is not the same quantity as the paper's tau (which is defined on an EasyCache-style accumulated-change rule), so the README section says to calibrate per model family rather than quoting the paper's constants.
• Not done: no notebook; the example needs a GPU and the image-reward package.

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.