Repository navigation
Fix: TCDScheduler advances the generator on the final step of a strength-truncated custom schedule - #15301
Fix: TCDScheduler advances the generator on the final step of a strength-truncated custom schedule#15301SIDDARTHAREDDY8 wants to merge 2 commits into
Conversation
…ted custom schedule Signed-off-by: SIDDARTHA REDDY <75976672+SIDDARTHAREDDY8@users.noreply.github.com>
1AbdullahButt
left a comment
There was a problem hiding this comment.
I ran the change on CPU against main with a small script that records, for each step() with eta=0.3, whether the generator state moved:
| schedule | main |
this PR |
|---|---|---|
num_inference_steps=4 |
[T, T, T, F] |
[T, T, T, F] |
num_inference_steps=4, strength=0.5 |
[T, T, T, F] |
[T, T, T, F] |
timesteps=[999, 499, 19], strength=0.5 → [19] |
[T] |
[F] |
So the fix does what #15288 asks, and the usual num_inference_steps + strength path is unchanged, since there num_inference_steps already matches len(self.timesteps).
One thing worth looking at here or in a follow-up: LCMScheduler.step has the same check (if self.step_index != self.num_inference_steps - 1:), and the same truncated custom schedule advances its generator on both main and this branch ([T] for [19]). TCD was derived from LCM, so applying the same one-line change there, with a matching test, would keep the two in line.
Small nit: the regression test asserts scheduler.num_inference_steps == 3, which pins the pre-truncation count. That is the current behaviour, but if someone later makes num_inference_steps follow the truncated schedule, this test would fail for a reason unrelated to the generator. Asserting only on timesteps and the RNG state would avoid that.
(I made this review with help from Claude Code. I ran the comparison script above; I did not run the full scheduler test suite.)
The regression test for huggingface#15288 already asserts on the truncated timesteps and the generator state; pinning num_inference_steps to its pre-truncation value adds nothing. Suggested in review on huggingface#15301. Signed-off-by: SIDDARTHA REDDY <75976672+SIDDARTHAREDDY8@users.noreply.github.com> Signed-off-by: Siddartha Reddy Chinthala <siddarthareddychinthala@gmail.com>
Fixes #15288.
Fix for #15288: in TCDScheduler.step, the final-step check now uses len(self.timesteps) - 1 (the actual, possibly strength-truncated schedule) instead of self.num_inference_steps - 1 (the pre-truncation count). One-line change in src/diffusers/schedulers/scheduling_tcd.py. Added regression test test_truncated_custom_timesteps_final_step_preserves_generator in tests/schedulers/test_scheduler_tcd.py covering truncated custom schedules with eta=0.3 across all three prediction types (epsilon, sample, v_prediction): single-step truncation [999,499,19] strength 0.5 (asserts timesteps [19], RNG state unchanged, next draw matches a fresh control generator, prev_sample == pred_noised_sample) and multi-step truncation [999,799,599,399,199] strength 0.6 (asserts only non-final steps consume randomness and subsequent draws match the control). Collision checks clean twice (before build and immediately before push): 0 PRs mention 15288 in body search and the issue timeline has no cross-referenced PRs; reporter twojian only offered a candidate fix in a comment, no PR exists. Branch pushed to SIDDARTHAREDDY8/diffusers oss-bhai-2026-11-15 at commit cbd4676; both files verified byte-identical on the branch via the contents API and the compare against upstream main shows exactly these 2 files, 1 commit ahead. NOTE FOR REVIEWER: push_branch.py could not be used as-is — the fork was stale (main HEAD a3e0b8e, Oct 20) so creating a branch at upstream HEAD 43e3002 returned 404, and merge-upstream sync is blocked because the PAT lacks workflow scope for .github/workflows/benchmark.yml. Pushed instead via a manual Git Data API flow parented on the fork's own main HEAD a3e0b8e. Safe because both changed files are byte-identical between fork HEAD and upstream HEAD (blob SHAs verified), so the PR diff is exactly the fix + test; the PR will show 148 commits behind, which is cosmetic. No PR opened (reviewer gates that).
Tests pass: Issue repro script confirmed the bug on unmodified main (43e3002): RNG state preserved False, next draw [-0.2934, -2.1788, ...] vs control [1.5410, -0.2934, ...], matching the issue log exactly. After the fix the repro passes: RNG state preserved True, next draw matches the control generator exactly, terminal prev_sample unchanged (0.218797...). Repo test suite tests/schedulers/test_scheduler_tcd.py: 35 passed, including the new regression test test_truncated_custom_timesteps_final_step_preserves_generator; the new test FAILS on unmodified code (verified via git stash) and passes with the fix. ruff check and ruff format clean on both changed files. Tests run in a venv with system-site-packages torch 2.14.0+cpu against a shallow clone at upstream HEAD 43e3002.