Describe the bug

Five Z-Image pipelines read the timestep off a device tensor inside the denoising loop, so every step blocks the CPU on a device-to-host copy:

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.

Same class of problem as #11696, #13404, #13406, #13461, and #13564. Found by following the examples/profiling guide, as #13401 asks.

On ZImageImg2ImgPipeline (Tongyi-MAI/Z-Image-Turbo, 1024x1024, 8 steps, one L40S) this is 1581 ms of blocked CPU per run. No wall-clock cost at this size, because the model is GPU-bound, but the loop cannot enter a CUDA graph while it syncs and the blocked CPU matters on a busier host.

Reproduction

Turn on torch.cuda.set_sync_debug_mode("error") from callback_on_step_end, after the one-off setup copies, and run any of the five. It raises RuntimeError: called a synchronizing CUDA operation at the timestep[0].item() line.

System Info

diffusers main, torch 2.9.1+cu128, one L40S.

Who can help?

@sayakpaul