Fixes #14907

What does this PR do?

The official CFG cutoff callbacks (SDCFGCutoffCallback, SDXLCFGCutoffCallback, SDXLControlnetCFGCutoffCallback, SD3CFGCutoffCallback) kept only the last row ([-1:]) of the [negative, conditional] batch when they disabled CFG. With num_images_per_prompt > 1 or several prompts the next step ran the UNet with batch_size latents and one embedding row and crashed (SD, SDXL, SDXL ControlNet), or, for SD3, silently conditioned every sample on the last prompt.

The callbacks now keep the conditional half of each tensor (x[-batch_size:] with batch_size = prompt_embeds.shape[0] // 2) and skip the slicing when CFG is off, since there is no negative batch to drop then. In guess mode the ControlNet image is not duplicated for CFG and [-batch_size:] keeps it whole.

Tests: test_cfg_cutoff_callback in the SD, SDXL, SDXL ControlNet and SD3 pipeline test files runs two prompts with a cutoff at step 1 and, through a second callback composed with MultiPipelineCallbacks, checks that prompt_embeds and latents keep the same batch size after the cutoff (plus the CFG-off and guess-mode cases). All four tests fail on main with assert 1 == 2.

Coordination: #14907 (issue filed with a reproduction and the proposed fix; no maintainer response yet, so this PR is opened early and I am happy to adjust or close it if you would prefer a different approach).

Test commands and results (CPU, macOS, torch 2.14.0):

$ python -m pytest tests/pipelines/stable_diffusion/test_stable_diffusion.py tests/pipelines/stable_diffusion_xl/test_stable_diffusion_xl.py tests/pipelines/controlnet/test_controlnet_sdxl.py tests/pipelines/stable_diffusion_3/test_pipeline_stable_diffusion_3.py -k test_cfg_cutoff_callback -q
5 passed, 434 deselected in 6.85s        # on main: 5 failed (assert 1 == 2)

$ python -m pytest tests/pipelines/stable_diffusion/test_stable_diffusion.py tests/pipelines/stable_diffusion_xl/test_stable_diffusion_xl.py tests/pipelines/controlnet/test_controlnet_sdxl.py tests/pipelines/stable_diffusion_3/test_pipeline_stable_diffusion_3.py -k "callback or interrupt" -q
21 passed, 418 deselected in 16.35s

$ make quality
ruff check: All checks passed!  ruff format --check: 2091 files already formatted  doc-builder / check_doc_toc / check_ai: clean
$ python utils/check_copies.py
clean

Self-review notes (final round, verdict: READY):

• Blocking issues: none.
• Non-blocking / deliberate calls:
  1. The if pipeline.do_classifier_free_guidance: guard is intentional: with CFG off there is no negative batch, and slicing would drop half of the real batch. All four target pipelines expose the property.
  2. SDXLControlnetCFGCutoffCallback with MultiControlNetModel: image is then a list of tensors and [-batch_size:] slices the list, exactly as [-1:] did before. That case was already broken and is left for a separate change.
  3. The SDXL ControlNet test stays on CPU because guess mode calls torch.logspace, which has no MPS kernel (same as test_controlnet_sdxl_guess).
  4. The small CheckBatchCallback test helper is repeated in the four test files; happy to move it to tests/pipelines/testing_utils if you prefer.
• Dead code: none added; the per-tensor local variables in the callbacks were removed.
• Docs: docs/source/en/using-diffusers/callback.md only shows the batch-size-1 usage, nothing there becomes stale.

Disclosure: found, reproduced and the draft fix tested locally with the help of an AI coding agent (Claude Code); reviewed before filing.

Before submitting

• Did you use an AI agent (Claude Code, Codex, Cursor, etc.) to help with this PR? If so:
  • Did you read the Coding with AI agents guide?
  • Did you run the self-review skill on the diff?
  • Did you share the final self-review notes in the PR description or a comment?
• Did you read the contributor guideline?
• Did you read our philosophy doc? (important for complex PRs)
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case. CFG cutoff callbacks keep a single embedding row, so they break for num_images_per_prompt > 1 or several prompts #14907
• Did you make sure to update the documentation with your changes? (no doc changes needed)
• Did you write any new necessary tests?
• Are you the author (or part of the team) of the model/pipeline (only applicable for model/pipeline related PRs)?

Who can review?

@asomoza @yiyixuxu

🤖 Generated with Claude Code