train_dreambooth_lora.py -- ValueError: Attempting to unscale FP16 gradients caused by "--validation_prompt" param.

[Xjmengnieer]

Describe the bug

Hello, when I directly run the training script provided in examples/dreambooth, with the following command, I encounter the “Attempting to unscale FP16 gradients” error:

`accelerate launch examples/dreambooth/train_dreambooth_lora.py \
  --mixed_precision="fp16" \
  --pretrained_model_name_or_path=$MODEL_NAME \
  --instance_data_dir=$INSTANCE_DIR \
  --output_dir=$OUTPUT_DIR \
  --instance_prompt="a photo of sks dog" \
  --resolution=512 \
  --train_batch_size=1 \
  --gradient_accumulation_steps=1 \
  --checkpointing_steps=100 \
  --learning_rate=1e-4 \
  --report_to="tensorboard" \
  --lr_scheduler="constant" \
  --lr_warmup_steps=0 \
  --max_train_steps=500 \
  --seed="0" \
  --validation_prompt="A photo of sks dog in a bucket" \
  --validation_epochs=50`

I then tried the solution proposed in PR #6554 and found that the following command does not trigger this issue:

> > `accelerate launch examples/dreambooth/train_dreambooth_lora.py \
> >   --mixed_precision="fp16" \
> >   --pretrained_model_name_or_path=$MODEL_NAME \
> >   --instance_data_dir=$INSTANCE_DIR \
> >   --output_dir=$OUTPUT_DIR \
> >   --instance_prompt="a photo of sks dog" \
> >   --resolution=512 \
> >   --train_batch_size=1 \
> >   --gradient_accumulation_steps=1 \
> >   --checkpointing_steps=100 \
> >   --learning_rate=1e-4 \
> >   --lr_scheduler="constant" \
> >   --lr_warmup_steps=0 \
> >   --max_train_steps=500 \
> >   --gradient_checkpointing \
> >   --seed="0" \
> >   --report_to="tensorboard"`
> 

I then tried to align the two commands and noticed that the issue is triggered specifically when the --validation_prompt argument is provided.

Does this indicate that there is still a bug related to validation in this script?

Reproduction

`accelerate launch examples/dreambooth/train_dreambooth_lora.py \
  --mixed_precision="fp16" \
  --pretrained_model_name_or_path=$MODEL_NAME \
  --instance_data_dir=$INSTANCE_DIR \
  --output_dir=$OUTPUT_DIR \
  --instance_prompt="a photo of sks dog" \
  --resolution=512 \
  --train_batch_size=1 \
  --gradient_accumulation_steps=1 \
  --checkpointing_steps=100 \
  --learning_rate=1e-4 \
  --report_to="tensorboard" \
  --lr_scheduler="constant" \
  --lr_warmup_steps=0 \
  --max_train_steps=500 \
  --seed="0" \
  --validation_prompt="A photo of sks dog in a bucket" \
  --validation_epochs=50`

Logs

System Info

• 🤗 Diffusers version: 0.37.0.dev0
• Platform: Linux-5.4.0-216-generic-x86_64-with-glibc2.35
• Running on Google Colab?: No
• Python version: 3.10.12
• PyTorch version (GPU?): 2.9.1+cu128 (True)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Huggingface_hub version: 1.4.1
• Transformers version: 5.1.0
• Accelerate version: 1.12.0
• PEFT version: 0.18.1
• Bitsandbytes version: not installed
• Safetensors version: 0.7.0
• xFormers version: not installed
• Accelerator: NVIDIA GeForce RTX 3060, 12288 MiB
• Using GPU in script?:
• Using distributed or parallel set-up in script?:

Who can help?

No response

[zhanglin2603]

Hey, I went over #13124 and can take this on.

• My read: this is primarily a core implementation task tied to "train_dreambooth_lora.py -- ValueError: Attempting to unscale FP16 gradients caused by "--".
• I will focus on these issue-specific points first: train_dreambooth_lora, valueerror, attempting.
• First I will reproduce and confirm expected behavior for the core implementation path.
• Then I will implement the fix with targeted tests and keep scope tight (good confidence).
• I will share a working PR update fast (24h for first patch, 48h for final polish) with verification notes.
  If this approach works for you, I can start now.

[zhanglin2603]

Quick follow-up on #13124: I am still available to work on this issue.
I can provide an initial implementation update within 24 hours and keep the scope tight.
If assignment changed, please let me know and I will avoid adding noise.