Remove per-step host syncs from the Z-Image denoising loops

[adrianrfreedman]

What does this PR do?

Removes the per-step device-to-host sync from the five Z-Image denoising loops that still have one. Split out of #14589 so each model family can be reviewed on its own. Part of #13401, same class of problem as #11696, #13404, #13406, #13461, and #13564.

The problem

Five Z-Image pipelines read the timestep off a device tensor inside the loop:

pipeline	read
z_image/pipeline_z_image_omni.py	timestep[0].item()
z_image/pipeline_z_image_controlnet.py	timestep[0].item()
z_image/pipeline_z_image_controlnet_inpaint.py	timestep[0].item()
z_image/pipeline_z_image_img2img.py	timestep[0].item()
z_image/pipeline_z_image_inpaint.py	timestep[0].item()

ZImagePipeline already solves this. It hoists _precomputed_t_norms = ((1000 - timesteps.float()) / 1000).tolist() out of the loop, guarded on cfg truncation. The other five were left behind, so this applies your existing pattern to them: same guard, same name, same expression.

Recomputing the normalisation in Python floats after tolist() is not equivalent, since the tensor arithmetic runs in the tensor's dtype, so hoisting the tensor expression is what keeps it bit-identical.

Quality

Unchanged. The normalisation is bit-identical in fp32, fp16, and bf16.

Measurements

ZImageImg2ImgPipeline, Tongyi-MAI/Z-Image-Turbo, 1024x1024, 8 steps, one L40S, each configuration run in both orders.

	main	this PR
sync check inside the loop	fails	passes
CPU blocked in syncs	1581 ms	110 ms
wall clock	4644.3 ms	4644.0 ms

No wall-clock change, eager or compiled. Expected, since the GPU does identical work and the model is GPU-bound at this size. What it buys is 1.5 s per run of unblocked CPU, which matters on a busier host, and a loop that no longer syncs at all.

I measured img2img. The other four take the identical change to a file I did benchmark.

Left alone

set_begin_index(0) is missing from omni, controlnet, and controlnet_inpaint, which is the separate sync from #11696. I have not touched it here, because img2img and inpaint can legitimately begin partway through the schedule and that deserves its own look.

Tests

tests/pipelines/z_image/: 112 passed, 39 skipped. No new test. The sync-debug test in #14589 pins the pattern, and adding one here would need the real checkpoint on a GPU runner. Say if you would rather have one.

Self-review

Ran the self-review skill over the diff. Nothing blocking.

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? Help us profile important pipelines and improve if needed #13401 asks for this, and it was discussed on Remove per-step host syncs from the CogVideoX denoising loop, enable regional compilation #14589.
• Did you make sure to update the documentation with your changes? No user-facing change.
• Did you write any new necessary tests? See above.
• Are you the author (or part of the team) of the model/pipeline (only applicable for model/pipeline related PRs)?

Who can review?

@sayakpaul