Consistency training fails to converge

**Consistency training fails to converge**
When I use examples/research_projects/consistency_training/train_cm_ct_unconditional.py to train the consistency model, I try to follow the experimental parameters of the original paper, but I cannot obtain normal generation results.

**Script**
```bash
python -m accelerate.commands.launch examples/research_projects/consistency_training/train_cm_ct_unconditional.py
--dataset_name="cifar10"
--dataset_image_column_name="img"
--output_dir="examples/research_projects/consistency_training/outputs_cifar10"
--resolution=32
--max_train_steps=400000
--dataloader_num_workers=8
--noise_precond_type="cm"
--input_precond_type="cm"
--train_batch_size=512
--gradient_accumulation_steps=2
--learning_rate=1e-04
--lr_scheduler="constant"
--lr_warmup_steps=0
--use_ema
--validation_steps=5000
--eval_batch_size=32
--checkpointing_steps=5000
--checkpoints_total_limit=10
--use_8bit_adam
--mixed_precision=fp16
--resume_from_checkpoint="latest"
```

**Generated results(step 400000)**

Start with examples/research_projects/consistency_training/train_cm_ct_unconditional.py and reproduce the reported command and parameters, including the step-400000 output. Compare the generated results with the original paper's expected behavior and trace the training configuration for the source of non-convergence. Done means the issue is explained and the script produces normal generation results or documents the required correction.

Scritto dal modello di indicizzazione a partire dal testo della issue.