High Memory and VRAM Usage When Processing Images

[LytsingStudio]

As stated in the title, processing a 4096x4096 image with the parameters below is extremely slow and requires over 48GB of VRAM and more than 128GB of system memory. Is this behavior expected?
I am using an RTX 4090 with 48GB of VRAM, but I cannot complete the test as it results in an OOM (Out of Memory) error.

parser.add_argument("--pretrained_model_name_or_path", type=str, default="./stable-diffusion-3-medium-diffusers/", required=False, help='path to the pretrained sd3')
parser.add_argument("--lora_dir", type=str, default="checkpoint/tsdsr", help='path to tsd-sr lora weights')
parser.add_argument("--embedding_dir", type=str, default="dataset/default/", help='path to prompt embeddings')
parser.add_argument("--output_dir", '-o', type=str, default="outputs/", help='path to save results')
parser.add_argument('--input_dir', '-i', type=str, default="inputs4/", required=False, help='path to the input image')

parser.add_argument("--rank", type=int, default=64, help='rank for transformer')
parser.add_argument("--rank_vae", type=int, default=64, help='rank for vae')

parser.add_argument("--is_use_tile", type=bool, default=True, help='whether to use tiled vae')
parser.add_argument("--vae_decoder_tiled_size", type=int, default=1024, help='tiled size for tiled vae decoder') 
parser.add_argument("--vae_encoder_tiled_size", type=int, default=1024, help='tiled size for tiled vae encoder') 
parser.add_argument("--latent_tiled_size", type=int, default=1024, help='tiled size for transformer latent')
parser.add_argument("--latent_tiled_overlap", type=int, default=32, help='tiled overlap for transformer latent')

parser.add_argument("--device", type=str, default="cuda")
parser.add_argument("--seed", type=int, default=42)
parser.add_argument("--upscale", type=int, default=4, help='upscale factor')
parser.add_argument("--process_size", type=int, default=512, help='process size for images')
parser.add_argument("--mixed_precision", type=str, choices=['fp16', 'fp32'], default="fp16")
parser.add_argument("--align_method", type=str, choices=['wavelet', 'adain', 'nofix'], default='wavelet', help='color alignment method')

[double8fun]

Using default tiling params consumes about 28GB VRAM on one L20 GPU.
--latent_tiled_size=1024 is meaningless since it is equal to 1024*8 pixel tiling, while the default 64 latent tile (512 pixel tile) is appropriate.

[Microtreei]

Here --upscale 4 is set so that the actual image processed is 16384x16384, the slow speed is expected, if you want to process 4096x4096, you need to set --upscale 1