lr_scheduler step bug in examples code #15087

Describe the bug

for example, the code in " diffusers--examples--dreambooth--train_dreambooth_lora_qwen_image.py": lr_scheduler = get_scheduler()这里前后的代码，The official diffusers training template for Accelerate has a mistake in scheduler creation. This causes lr scheduler total steps / warmup steps to be N times larger than real global update steps, learning rate never decays normally. Also there is a ordering bug:
The script calculates training steps & creates lr_scheduler before accelerator.prepare(train_dataloader).
At this point len(train_dataloader) is the non-distributed value (not sharded). After prepare, dataloader length changes, but scheduler has already been initialized with wrong step count.

This is a potential bug that is difficult to detect in schedule constant mode.

Reproduction

1. Use diffusers official training script + Accelerator
2. 8 GPU, per_device_batch_size=1, gradient_accumulation_steps=4
3. Before prepare(): len(train_dataloader)=3000
4. Create scheduler using this len, multiplied by accelerator.num_processes
5. After prepare(), real len(train_dataloader)=375, but scheduler already built.

Logs

System Info

Run diffusers-cli env and paste output here

• diffusers version: 0.36
• accelerate version: 1.10
• OS: ubuntu

Who can help?

No response