Use configured timesteps in DDPM parallel batch steps

[Robertboy18]

What does this PR do?

I was comparing the DDPM parallel batch path with the regular scheduler step and noticed that batch_step_no_noise() still used a fixed training-step stride. This is the remaining batched case from #9261 (Fixes#9261); the scalar path was already fixed in #9384.

For linspace, trailing, and custom schedules, that could choose different alpha coefficients from the actual next configured timestep. This now follows scheduler.timesteps and adds parity tests against the scalar deterministic step :)

I also have a small TorchLean formulation of the schedule-successor invariant and can attach it if useful.

Tests

• pytest tests/schedulers/test_scheduler_ddpm_parallel.py -q: 40 passed
• make modified_only_fixup
• make fix-copies

Self-review

• Blocking issues: none.
• Non-blocking issues: none.
• Dead code: none.
• Verdict: READY.

Before submitting

• I used an AI-assisted workflow.
  • I read the Coding with AI agents guide.
  • I ran the repository self-review against the current diff.
  • The final self-review notes are included above.
• I read the contributor guideline and philosophy doc.
• The existing maintainer-acknowledged discussion is linked above.
• No documentation update is needed because this restores parity with the configured scheduler semantics.
• I added regression tests.

[Copilot]

🟡 Changes recommended

The new schedule lookup can silently choose an incorrect prev_t for unexpected timesteps (no-match case), and the new test’s mock should be device/dtype/shape-aware to avoid brittleness.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR fixes a schedule-parity bug in DDPMParallelScheduler.batch_step_no_noise() by deriving prev_t from the configured scheduler.timesteps (including linspace, trailing, and explicit custom timesteps) rather than using a fixed training-stride heuristic, and adds a regression test that checks batched vs scalar-step parity under deterministic (zero-noise) conditions.

Changes:

• Update batch_step_no_noise() to compute prev_t via the configured timestep schedule when running with inference/custom timesteps.
• Add a parity test ensuring batch_step_no_noise() matches repeated scalar step() across multiple timestep schedules.
• Add test-time mocking to force deterministic (zero) noise in the scalar path for comparison.

File summaries

File	Description
src/diffusers/schedulers/scheduling_ddpm_parallel.py	Updates batch_step_no_noise() to follow the configured timestep schedule when computing the previous timestep.
tests/schedulers/test_scheduler_ddpm_parallel.py	Adds regression coverage validating batched no-noise stepping matches scalar stepping across multiple schedules.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 2
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[github-actions]

Hi @Robertboy18, 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.