Inference error in Flux with quantization & LoRA applied, and a bug of Quanto with Zero GPU spaces

### Describe the bug

Merry Christmas.🎅

The inference of the Flux model with LoRA applied and the inference of the quantized Flux model work fine on their own, but when combined, they often result in an error.
```
RuntimeError('Only Tensors of floating point and complex dtype can require gradients')
```
I don't know which library or environment is the main cause of this error, but I'm posting it here because it's easy to confirm the bug with the combination of Diffusers and Flux.
In the following demo, I'm using the pip version, but the same error occurs with the github version of Diffusers and PEFT.
The code is written in a roundabout way to avoid bugs in the Zero GPU space.

## Demo Space for error reproduction (for Huggingface ZeroGPU)
https://huggingface.co/spaces/John6666/diffusers_lora_error_test1

### P.S.
The demo that was working a few minutes ago has stopped working due to a new error...
```
RuntimeError('NVML_SUCCESS == r INTERNAL ASSERT FAILED at "../c10/cuda/CUDACachingAllocator.cpp":838, please report a bug to PyTorch. ')
```

## Extra: Zero GPU space and Quanto are too incompatible
In this demo, **just by writing the following line**, the inference in Zero GPU space **will always crash**. I think CUDA is being called within Quanto.🥶
```py
from optimum.quanto import freeze, qfloat8, quantize
```
```
File "/usr/local/lib/python3.10/site-packages/spaces/zero/wrappers.py", line 214, in gradio_handler
raise res.value
RuntimeError: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method
```

### Reproduction

```py
# for Huggingface ZeroGPU Spaces
import spaces
import gradio as gr
import torch
from huggingface_hub import hf_hub_download
from diffusers import FluxPipeline, FluxTransformer2DModel, GGUFQuantizationConfig, BitsAndBytesConfig
import os
import subprocess
#subprocess.run("pip list", shell=True)
#subprocess.run("diffusers-cli env", shell=True)
#from optimum.quanto import freeze, qfloat8, quantize

HF_TOKEN = os.getenv("HF_TOKEN", "")
device = "cuda" if torch.cuda.is_available() else "cpu"
flux_repo = "multimodalart/FLUX.1-dev2pro-full"
ckpt_path = "https://huggingface.co/city96/FLUX.1-dev-gguf/blob/main/flux1-dev-Q2_K.gguf"
transformer_gguf = FluxTransformer2DModel.from_single_file(ckpt_path, subfolder="transformer", quantization_config=GGUFQuantizationConfig(compute_dtype=torch.bfloat16),
torch_dtype=torch.bfloat16, config=flux_repo, token=HF_TOKEN)
transformer = FluxTransformer2DModel.from_pretrained(flux_repo, subfolder="transformer", torch_dtype=torch.bfloat16, token=HF_TOKEN)
nf4_quantization_config = BitsAndBytesConfig(load_in_4bit=True)
transformer_nf4 = FluxTransformer2DModel.from_pretrained(flux_repo, subfolder="transformer", quantization_config=nf4_quantization_config,
torch_dtype=torch.bfloat16, token=HF_TOKEN)
pipe = FluxPipeline.from_pretrained(flux_repo, transformer=transformer, torch_dtype=torch.bfloat16, token=HF_TOKEN)
hyper_sd_lora = hf_hub_download("ByteDance/Hyper-SD", "Hyper-FLUX.1-dev-8steps-lora.safetensors")

@spaces.GPU(duration=70)
def infer(prompt: str, mode: str, is_lora: bool, progress=gr.Progress(track_tqdm=True)):
global pipe
try:
pipe.unload_lora_weights()
if mode == "Default": pipe.transformer = transformer
elif mode == "GGUF": pipe.transformer = transformer_gguf
elif mode == "NF4": pipe.transformer = transformer_nf4
if is_lora:
pipe.load_lora_weights(hyper_sd_lora, adapter_name="hyper-sd")
pipe.set_adapters(["hyper-sd"], adapter_weights=[0.125])
steps = 8
else: steps = 28
pipe.to(device)
image = pipe(prompt, generator=torch.manual_seed(0), num_inference_steps=steps).images[0]
pipe.to("cpu")
return image
except Exception as e:
raise gr.Error(e)

with gr.Blocks() as demo:
with gr.Row():
with gr.Column():
prompt = gr.Textbox(label="Prompt", value="A cat holding a sign that says hello world", lines=1)
mode = gr.Radio(label="Mode", choices=["Default", "GGUF", "NF4"], value="Default")
is_lora = gr.Checkbox(label="Enable LoRA", value=True)
gen_btn = gr.Button("Generate Image")
with gr.Column():
result = gr.Image(label="Result Image")

gen_btn.click(infer, [prompt, mode, is_lora], [result])

demo.launch()
```
```
huggingface_hub
torch
diffusers
peft
transformers
accelerate
numpy<2
gguf
bitsandbytes
optimum-quanto
```

### Logs

_No response_

### System Info

- 🤗 Diffusers version: 0.32.0
- Platform: Linux-5.10.228-219.884.amzn2.x86_64-x86_64-with-glibc2.36
- Running on Google Colab?: No
- Python version: 3.10.13
- PyTorch version (GPU?): 2.4.0+cu121 (False)
- Flax version (CPU?/GPU?/TPU?): not installed (NA)
- Jax version: not installed
- JaxLib version: not installed
- Huggingface_hub version: 0.25.1
- Transformers version: 4.47.1
- Accelerate version: 1.2.1
- PEFT version: 0.14.0
- Bitsandbytes version: 0.45.0
- Safetensors version: 0.4.5
- xFormers version: not installed
- Accelerator: Failed to initialize NVML: Unknown Error
- Using GPU in script?:
- Using distributed or parallel set-up in script?:

### Who can help?

_No response_

Start with the linked Hugging Face ZeroGPU demo and its reproduction script, focusing on FluxPipeline, the quantized transformer modes, and LoRA loading. Compare the Default, GGUF, and NF4 paths with and without LoRA, then separately test the optimum.quanto import; done requires isolating the failing combinations and defining the expected behavior for each.

Scritto dal modello di indicizzazione a partire dal testo della issue.