feat(example): add SDXL MXFP8 RTN workflow

[changwangss]

Type of Change

feature

Description

Add a standalone Stable Diffusion XL MXFP8 example under stable_diffusion/mxfp8. The example provides:

• calibration-free pure RTN quantization with iters=0 and disable_opt_rtn=True
• BF16 and MXFP8 image generation
• balanced multi-GPU COCO2014 dataset splitting without dropping samples
• CLIP, CLIP-IQA, and ImageReward evaluation
• documented SDXL COCO2014 BF16 and MXFP8 RTN results

The workflow targets stabilityai/stable-diffusion-xl-base-1.0 and keeps the existing Flux example unchanged.

Expected Behavior & Potential Risk

Users can produce a complete fake-quantized Diffusers SDXL pipeline with MXFP8 pure RTN, then run single- or multi-GPU evaluation. No calibration dataset or tuning samples are required for quantization.

The change only adds a new example directory and does not modify library code or existing examples. Full model quantization is GPU- and resource-intensive.

How has this PR been tested?

• Python syntax compilation for main.py and dataset_split.py
• Ruff checks for both Python files
• Bash syntax checks for run_quant.sh and run_benchmark.sh
• Balanced dataset split smoke test preserving every input row
• AutoRoundConfig verification confirming scheme=MXFP8, iters=0, and disable_opt_rtn=True
• git diff --check

Dependency Change?

No library dependency changes. The new example includes its own requirements.txt.