[WIP] Add TRT as a Backend

[ishan-modi]

What does this PR do?

WIP, aimed at adding new backend for quantization #11032

Depends on this

# !pip install "git+https://github.com/ishan-modi/diffusers.git@add-trtquant-backend"
# !pip install "git+https://github.com/ishan-modi/TensorRT-Model-Optimizer.git@fixes-hfargs-order#egg=nvidia-modelopt[torch, onnx]"

import torch
from diffusers import SanaTransformer2DModel

checkpoint = "Efficient-Large-Model/Sana_600M_1024px_diffusers"

# quant_config = {"quant_type": "INT4_WO", "quant_method": "modelopt", "modules_to_not_convert": ["conv"], "block_quantize": True}
quant_config = {"quant_type": "FP8_WO", "quant_method": "modelopt", "modules_to_not_convert": ["conv"]}
model = SanaTransformer2DModel.from_pretrained(checkpoint, subfolder="transformer", quantization_config=quant_config, torch_dtype=torch.bfloat16).to('cuda')

hidden = torch.randn((2,32,32,32), dtype=torch.bfloat16).to('cuda')
encoder = torch.randn((2,10,300,2304), dtype=torch.bfloat16).to('cuda')
timestep = torch.Tensor([0,0]).to('cuda')
with torch.no_grad():
    output = model(hidden, encoder, timestep)

[ishan-modi]

@sayakpaul, would you mind giving a quick look and suggestions