Add a (rolling) KV cache for Wan models to enable autoregressive rollouts ("SelfForcing").#13681

Draft

gueraf wants to merge 19 commits intohuggingface:mainfrom

gueraf:wan-rolling-kv-cache

Commits on May 5, 2026

Convert rolling KV cache tests from unittest.TestCase to plain pytest

gueraf
and
claude
committed
Trim rolling KV cache tests: remove low-value cases, localize helpers

gueraf
and
claude
committed
Consolidate rolling KV cache tests into single class

gueraf
and
claude
committed
Clean up TestWanRollingKVCache for readability

gueraf
and
claude
committed
Test via public interface: remove private helper imports and trim/slice unit tests

gueraf
and
claude
committed
Fix docstring example: drop explicit window_size=8000 (default is already -1)

gueraf
and
claude
committed
Assert actual key content in rolling KV cache tests; add docstrings

gueraf
and
claude
committed
Rewrite cache tests: snapshot/diff pattern, rolled-out loops, assert helpers

gueraf
and
claude
committed
Cover full token range in each test with both _assert_changed and _assert_unchanged

gueraf
and
claude
committed
Rename _run -> _denoise_chunk

gueraf
and
claude
committed
Refactor assert helpers to take from_token/to_token ints; clean up call sites

gueraf
and
claude
committed
Remove bogus _assert_changed from append test; only assert prefix preservation

gueraf
and
claude
committed
Restructure cache tests around (mode, window_size); add overwrite chunk-loop test

gueraf
and
claude
committed
Narrow rolling KV cache API: append + overwrite_end (drop arbitrary offsets)

gueraf
and
claude
committed
Test windowed cache with multiple chunks; per-test caches; simplify drop-last helper

gueraf
and
claude
committed