运行Diffusers的train_text_to_image_lora.py时报错：Attempting to unscale FP16 gradients.

项目地址：

GitHub - huggingface/diffusers: 🤗 Diffusers: State-of-the-art diffusion models for image and audio generation in PyTorch🤗 Diffusers: State-of-the-art diffusion models for image and audio generation in PyTorch - huggingface/diffusershttps://github.com/huggingface/diffusers/tree/main在运行Lora微调案例：train_text_to_image_lora.py时出现如下报错：

Traceback (most recent call last):
  File "/root/LoRA/diffusers/examples/text_to_image/train_text_to_image_lora.py", line 969, in <module>
    main()
  File "/root/LoRA/diffusers/examples/text_to_image/train_text_to_image_lora.py", line 796, in main
    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
  File "/root/miniconda3/envs/myconda/lib/python3.10/site-packages/accelerate/accelerator.py", line 2101, in clip_grad_norm_
    self.unscale_gradients()
  File "/root/miniconda3/envs/myconda/lib/python3.10/site-packages/accelerate/accelerator.py", line 2064, in unscale_gradients
    self.scaler.unscale_(opt)
  File "/root/miniconda3/envs/myconda/lib/python3.10/site-packages/torch/cuda/amp/grad_scaler.py", line 284, in unscale_
    optimizer_state["found_inf_per_device"] = self._unscale_grads_(optimizer, inv_scale, found_inf, False)
  File "/root/miniconda3/envs/myconda/lib/python3.10/site-packages/torch/cuda/amp/grad_scaler.py", line 212, in _unscale_grads_
    raise ValueError("Attempting to unscale FP16 gradients.")

issues里面提供了各种方法，但是尝试了一遍都没有解决(也可能是我没理解)

最后，误打误撞：

解决办法：把FP16改成了bf16解决了（注意在accelerate config和训练脚本里面都要设置成bf16）

accelerate launch --mixed_precision="bf16"  train_text_to_image_lora.py \ # 这里原来是fp16
  --pretrained_model_name_or_path=$MODEL_NAME \
  --dataset_name=$DATASET_NAME \
  --dataloader_num_workers=8 \
  --resolution=512 \
  --center_crop \
  --random_flip \
  --train_batch_size=1 \
  --gradient_accumulation_steps=4 \
  --max_train_steps=15000 \
  --learning_rate=1e-04 \
  --max_grad_norm=1 \
  --lr_scheduler="cosine" \
  --lr_warmup_steps=0 \
  --output_dir=${OUTPUT_DIR} \
  --push_to_hub \
  --hub_model_id=${HUB_MODEL_ID} \
  --report_to=wandb \
  --checkpointing_steps=500 \
  --validation_prompt="A pokemon with blue eyes." \
  --seed=1337