[training] fix training resuming problem for fp16 (SD LoRA DreamBooth)

[a-r-r-o-w]

What does this PR do?

Part of #6552.

I'm yet to test it on a training run. I think Sayak mentioned he'll be opening a follow-up PR to add a utility function that removes the duplicated code.

First run:

CUDA_VISIBLE_DEVICES=0 accelerate launch train_dreambooth_lora.py \
  --pretrained_model_name_or_path=runwayml/stable-diffusion-v1-5 \
  --instance_data_dir="dog" \
  --mixed_precision="fp16" \
  --instance_prompt="a photo of sks dog" \
  --output_dir="lora-dog-sd" \
  --resolution=512 \
  --train_batch_size=1 \
  --gradient_accumulation_steps=1 --gradient_checkpointing \
  --learning_rate=1e-4 \
  --lr_scheduler="constant" \
  --lr_warmup_steps=0 \
  --max_train_steps=4 --checkpointing_steps=2 --checkpoints_total_limit=2 \
  --use_8bit_adam \
  --seed="42"

Resume training:

CUDA_VISIBLE_DEVICES=0 accelerate launch train_dreambooth_lora.py \
  --pretrained_model_name_or_path=runwayml/stable-diffusion-v1-5 \
  --instance_data_dir="dog" \
  --mixed_precision="fp16" \
  --instance_prompt="a photo of sks dog" \
  --output_dir="lora-dog-sd" \
  --resolution=512 \
  --train_batch_size=1 \
  --gradient_accumulation_steps=1 --gradient_checkpointing \
  --learning_rate=1e-4 \
  --lr_scheduler="constant" \
  --lr_warmup_steps=0 \
  --max_train_steps=4 --checkpointing_steps=2 --checkpoints_total_limit=2 \
  --resume_from_checkpoint="latest" \
  --use_8bit_adam \
  --seed="42"

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@sayakpaul

[sayakpaul]

Super fast! I have updated the description of #6552. It should be clear now.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[sayakpaul]

@a-r-r-o-w let me know if I can test it.

[a-r-r-o-w]

Hi Sayak, this is ready for testing I believe, and seems to be working well. Here are my logs:

[sayakpaul]

Just tested. Seems to be working well! Thank you!