Update README.md

Remove onnxruntime

Revert to PyTorch INT8 (ONNX export produces NaN)

Reduce ORT memory: disable prepacking, basic optimization, 1 thread

Fix: unsqueeze prompt embeddings

Add onnxruntime dependency

Switch to ONNX INT8 DiT

fixes: cudagc guard, rm conditioner.py, turbo depth colormap, proper normal viz, compact UI, example

fix: .weight.dtype crashes on INT8 quantized Linear, use .float()

fix: use FP32 dtype on CPU (no bfloat16 autocast with INT8 model)

fix: load full torch.save model directly (no FP32 construction, mmap)

fix OOM: use mmap loading + assign (no memory copy)

fix: eager model load at startup (lazy load causes SSE timeout)

fix: move demo to module level (Gradio SDK needs it, not inside main())

add example images for lazy-loaded examples

switch to Gradio SDK (no Docker needed for pure Python)

fix: PYTHONUNBUFFERED for Docker log visibility

use pre-quantized INT8 model (no FP8 casting, no LoRA merge at runtime)

fix: diffusers dep, layer-by-layer FP8->FP32 cast, LoRA merge in FP32, INT8 quant

FE2E depth+normal CPU Space: FP8 dynamic INT8, single denoise