(sd3 branch Flux LoRA Training) "RuntimeError: only Tensors of floating point dtype can require gradients" occured while training Flux LoRA

Using sd3 Branch code training Flux LoRA the following error occured.
```python
/opt/conda/lib/python3.10/site-packages/torch/utils/checkpoint.py:61: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
warnings.warn(
Traceback (most recent call last):
File "/root/private_data/sd-scripts/flux_train_network.py", line 411, in
trainer.train(args)
File "/root/private_data/sd-scripts/train_network.py", line 1100, in train
noise_pred, target, timesteps, huber_c, weighting = self.get_noise_pred_and_target(
File "/root/private_data/sd-scripts/flux_train_network.py", line 300, in get_noise_pred_and_target
t.requires_grad_(True)
RuntimeError: only Tensors of floating point dtype can require gradients
```
My config (With a 64GB VRAM GPU):
```python
sdpa = True
lowram = False
highvram = True
cache_latents = True
cache_latents_to_disk = False
cache_text_encoder_outputs = False
cache_text_encoder_outputs_to_disk = False
no_half_vae = True
mixed_precision = "bf16"
save_precision = "bf16"
full_fp16 = False
full_bf16 = False
fp8_base = False
```

I'm not sure is this reason cause this issue.

I made a temporary fix #1508

Please check it.

Start with flux_train_network.py around line 300, where the traceback reaches t.requires_grad_(True), and compare the behavior with the temporary fix referenced in issue #1508. Reproduce using the posted Flux LoRA configuration and confirm that training proceeds without the reported gradient dtype RuntimeError.

Written by the indexing model from the issue text.