Quantizing SDXL for MLPerf
The purpose of this document is to allow other users to quantize SDXL in various different ways, corresponding to AMD's 2024-07 MLPerf submission.
Environmental Setup
The environment can be set up as follows:
conda env create -n quant_sdxl -f env/brv_sdxl_mlperf_minimal.yml # Use env/brv_sdxl_mlperf_rocm_minimal.yml is using a ROCm compatible GPU
conda activate quant_sdxl
Fetching and Pre-Processing Datasets
Information on fetching and pre-processing the data can be found here. Note, the latents generated here will be used in the next step.
Quantize Int8 / FP16 Model
python quant_sdxl.py --model stabilityai/stable-diffusion-xl-base-1.0 --device <device> --calibration-prompt-path ./captions.tsv --checkpoint-name unet.ckpt --path-to-latents <path/to/latents/latents.pt> --guidance-scale 7.5 --exclude-blacklist-act-eq [--path-to-coco <path/to/coco> --validation-prompts 5000]
Note, you can optionally validate on the MLPerf validation set, but be aware this will take ~48hrs.
Quantize Int8 / FP8 Model
The Int8 / FP8 model can be quantized as follows:
python quant_sdxl.py --model stabilityai/stable-diffusion-xl-base-1.0 --device <device> --calibration-prompt-path ./captions.tsv --checkpoint-name unet.ckpt --path-to-latents <path/to/latents/latents.pt> --guidance-scale 7.5 --quantize-sdp --exclude-blacklist-act-eq [--path-to-coco <path/to/coco> --validation-prompts 5000]