train_text_to_image_lora.py raise ValueError("Attempting to unscale FP16 gradients.")

[billvsme]

Describe the bug

When looking at the examples/text_to_image documentation, I experimented with the train_text_to_image_lora.py following the examples in the documentation. But I found that the run with raise ValueError("Attempting to unscale FP16 gradients.") error.

I found that the cause of the error may be related to this code. Here use args.mixed_precision to determine whether to convert Lora's parameters to float32, but args.mixed_precision default value is None, according to the example in README, the mixed_precision of accelerate is set, and it is not set args.mixed_ precision, so it causes "Attempting to unscale FP16 gradients." error.

diffusers/examples/text_to_image/train_text_to_image_lora.py

Lines 468 to 472 in 1fff527

	if args.mixed_precision == "fp16":
	for param in unet.parameters():
	# only upcast trainable parameters (LoRA) into fp32
	if param.requires_grad:
	param.data = param.to(torch.float32)

It might be a better choice to change this to use accelerator.mixed_precision

Reproduction

cd diffusers/examples/text_to_image/

accelerate launch --mixed_precision="fp16" train_text_to_image_lora.py \
  --pretrained_model_name_or_path="runwayml/stable-diffusion-v1-5" \
  --dataset_name="lambdalabs/pokemon-blip-captions" --caption_column="text" \
  --resolution=512 --random_flip \
  --train_batch_size=1 \
  --num_train_epochs=100 --checkpointing_steps=5000 \
  --learning_rate=1e-04 --lr_scheduler="constant" --lr_warmup_steps=0 \
  --seed=42 \
  --output_dir="sd-pokemon-model-lora" \
  --validation_prompt="cute dragon creature"

Logs

Steps:   0%|                                          | 0/20900 [00:00<?, ?it/s]Traceback (most recent call last):
  File "/content/diffusers/examples/text_to_image/train_text_to_image_lora.py", line 945, in <module>
Traceback (most recent call last):
  File "/content/diffusers/examples/text_to_image/train_text_to_image_lora.py", line 945, in <module>
    main()
  File "/content/diffusers/examples/text_to_image/train_text_to_image_lora.py", line 774, in main
    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
  File "/home/billvsme/venv/train/lib/python3.10/site-packages/accelerate/accelerator.py", line 2040, in clip_grad_norm_
    main()
  File "/content/diffusers/examples/text_to_image/train_text_to_image_lora.py", line 774, in main
Traceback (most recent call last):
  File "/content/diffusers/examples/text_to_image/train_text_to_image_lora.py", line 945, in <module>
    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
  File "/home/billvsme/venv/train/lib/python3.10/site-packages/accelerate/accelerator.py", line 2040, in clip_grad_norm_
    self.unscale_gradients()
  File "/home/billvsme/venv/train/lib/python3.10/site-packages/accelerate/accelerator.py", line 2003, in unscale_gradients
    main()
  File "/content/diffusers/examples/text_to_image/train_text_to_image_lora.py", line 774, in main
    self.unscale_gradients()
    self.scaler.unscale_(opt)
  File "/home/billvsme/venv/train/lib/python3.10/site-packages/accelerate/accelerator.py", line 2003, in unscale_gradients
  File "/home/billvsme/venv/train/lib/python3.10/site-packages/torch/cuda/amp/grad_scaler.py", line 307, in unscale_
    optimizer_state["found_inf_per_device"] = self._unscale_grads_(
  File "/home/billvsme/venv/train/lib/python3.10/site-packages/torch/cuda/amp/grad_scaler.py", line 229, in _unscale_grads_
    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
  File "/home/billvsme/venv/train/lib/python3.10/site-packages/accelerate/accelerator.py", line 2040, in clip_grad_norm_
    raise ValueError("Attempting to unscale FP16 gradients.")
ValueError: Attempting to unscale FP16 gradients.
    self.scaler.unscale_(opt)
  File "/home/billvsme/venv/train/lib/python3.10/site-packages/torch/cuda/amp/grad_scaler.py", line 307, in unscale_
    optimizer_state["found_inf_per_device"] = self._unscale_grads_(
  File "/home/billvsme/venv/train/lib/python3.10/site-packages/torch/cuda/amp/grad_scaler.py", line 229, in _unscale_grads_
    self.unscale_gradients()
  File "/home/billvsme/venv/train/lib/python3.10/site-packages/accelerate/accelerator.py", line 2003, in unscale_gradients
    raise ValueError("Attempting to unscale FP16 gradients.")

System Info

• diffusers version: 0.25.0.dev0
• Platform: Linux-6.2.0-39-generic-x86_64-with-glibc2.35
• Python version: 3.10.13
• PyTorch version (GPU?): 2.1.2+cu121 (True)
• Huggingface_hub version: 0.19.4
• Transformers version: 4.36.2
• Accelerate version: 0.25.0
• xFormers version: 0.0.22.post7
• Using GPU in script?:
• Using distributed or parallel set-up in script?:

Who can help?

@sayakpaul

[sayakpaul]

A better way would be to assign args.mixed_precision from accelerator.mixed_precision.

However, when you initialize an Accelerator object you pass the value from args.mixed_precision itself:

diffusers/examples/text_to_image/train_text_to_image_lora.py

Line 385 in 1fff527

mixed_precision=args.mixed_precision,

So, passing mixed_precision to your CLI args is recommended.

[billvsme]

@sayakpaul 👌，thanks

But I found one that was different from train_text_to_image.py and train_text_to_image_lora.py, train_text_to_image_lora.py didn't reassign the args.mixed_precision. In this way, if you specify accelerate launch --mixed_precision="fp16" in the accelerator, you need to add the same --mixed_precision="fp16" to the CLI args . Only in this way will there be no error, like is

accelerate launch --mixed_precision="fp16" train_text_to_image_lora.py \
  --mixed_precision="fp16" \
  ......

train_text_to_image.py:

diffusers/examples/text_to_image/train_text_to_image.py

Lines 811 to 816 in 1fff527

	if accelerator.mixed_precision == "fp16":
	weight_dtype = torch.float16
	args.mixed_precision = accelerator.mixed_precision
	elif accelerator.mixed_precision == "bf16":
	weight_dtype = torch.bfloat16
	args.mixed_precision = accelerator.mixed_precision

train_text_to_image_lora.py:

diffusers/examples/text_to_image/train_text_to_image_lora.py

Lines 444 to 448 in 1fff527

	weight_dtype = torch.float32
	if accelerator.mixed_precision == "fp16":
	weight_dtype = torch.float16
	elif accelerator.mixed_precision == "bf16":
	weight_dtype = torch.bfloat16

[billvsme]

Maybe the example in the docs needs to be updated

https://github.com/huggingface/diffusers/tree/main/examples/text_to_image

[sayakpaul]

Should be fixed with: #6388. Could you pull the changes and try again? :)

[AfrinaVT]

Hi @sayakpaul , The problem with running train_text_to_image_lora.py still persists for me. I have pulled the latest changes from the GitHub repo.

[sayakpaul]

Could you maybe refer to #6552 and open a PR?

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[yiyixuxu]

can we close this one now?

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.