Skip to content

Latest commit

 

History

History
 
 

stable-diffusion-xl

README.md

Stable Diffusion XL for Intel® Gaudi® MLPerf™ Inference Submission

This directory provides instructions to reproduce Intel Gaudi's results for MLPerf™ inference submission.
MLPerf™ is a trademark and service mark of MLCommons Association in the United States and other countries.
All rights reserved. Unauthorized use is strictly prohibited.

Setup

Please follow the instructions provided in the Intel Gaudi Installation Guide to set up the environment.

Prepare Intel-HabanaLabs MLPerf Inference Container

mkdir -p /path/to/Intel-HabanaLabs
export INTEL_HABANALABS_DIR=/path/to/Intel-HabanaLabs

This README is located in code directory corresponding to Intel-HabanaLabs submission. Download the whole code folder along with all subfolders and copy it under $INTEL_HABANALABS_DIR:

docker run --privileged --security-opt seccomp=unconfined \
           --name mlperf-habana -td                \
           -v /dev:/dev                            \
           --device=/dev:/dev                      \
           -v /sys/kernel/debug:/sys/kernel/debug  \
           -v /tmp:/tmp                            \
           -v $INTEL_HABANALABS_DIR:/root/Intel-HabanaLabs/  \
           --cap-add=sys_nice --cap-add=SYS_PTRACE \
           --user root --workdir=/root --net=host  \
           --ulimit memlock=-1:-1 vault.habana.ai/gaudi-docker-mlperf/ver4.0/pytorch-installer-2.1.1:1.14.98-33
docker exec -it mlperf-habana bash

Download Checkpoint

mkdir -p /mnt/weka/data/mlperf_inference/stable-diffusion-xl/stable_diffusion_fp32
pushd /mnt/weka/data/mlperf_inference/stable-diffusion-xl
curl https://rclone.org/install.sh | bash
rclone config create mlc-inference s3 provider=Cloudflare access_key_id=f65ba5eef400db161ea49967de89f47b secret_access_key=fbea333914c292b854f14d3fe232bad6c5407bf0ab1bebf78833c2b359bdfd2b endpoint=https://c2686074cb2caf5cbaf6d134bdba8b47.r2.cloudflarestorage.com
rclone copy mlc-inference:mlcommons-inference-wg-public/stable_diffusion_fp32 ./stable_diffusion_fp32 -P
popd

Alternatively, the required checkpoints/datasets can be downloaded offline and copied to the required path before running Docker run command.

Download Statistics File for Calculating FID

To download statistics file:

pushd /root/Intel-HabanaLabs/code/stable-diffusion-xl/stable-diffusion-xl/tools
wget -L https://github.com/mlcommons/inference/raw/master/text_to_image/tools/val2014.npz
popd

Download Dataset (Optional)

To download dataset, run the below command:

pushd /root/Intel-HabanaLabs/code/stable-diffusion-xl/stable-diffusion-xl/tools
./download-coco-2014.sh -n 1
popd

build_mlperf_inference covers the same functionality.

Reproduce Results

Get Started

Install the requirements and build the latest loadgen:

cd /root/Intel-HabanaLabs/code
source stable-diffusion-xl/functions.sh
pip install -r stable-diffusion-xl/stable-diffusion-xl/requirements.txt
build_mlperf_inference

Generate Results

To generate full submission results, run the following command:

build_mlperf_inference --output-dir <path_to_output_dir> --submission sd-xl-fp8

The command produces results from accuracy and performance runs for both Offline and Server scenarios. Logs can be found under /path_to_output_dir/logs/model/, e.g. /results/logs/sd-xl-fp8/

To generate results for Offline and Server scenarios separately, run the following commands:

source stable-diffusion-xl/functions.sh
build_mlperf_inference --output-dir <path_to_output_dir> --submission sd-xl-fp8_Offline
source stable-diffusion-xl/functions.sh
build_mlperf_inference --output-dir <path_to_output_dir> --submission sd-xl-fp8_Server

Logs can be found under /path_to_output_dir/logs/model/scenario/, e.g. /results/logs/sd-xl-fp8/Offline/

To generate results for accuracy and performance separately, add --mode flag as in one of the following commands:

source stable-diffusion-xl/functions.sh
build_mlperf_inference --output-dir <path_to_output_dir> --submission sd-xl-fp8_Offline --mode acc
source stable-diffusion-xl/functions.sh
build_mlperf_inference --output-dir <path_to_output_dir> --submission sd-xl-fp8_Offline --mode perf

Logs can be found under /path_to_output_dir/logs/model/scenario/mode/, e.g. /results/logs/sd-xl-fp8/Offline/accuracy/

Calibration Steps (Optional)

The below command recreates the measurements on the calibration dataset which we later use to determine the scales:

pushd /root/Intel-HabanaLabs/code/stable-diffusion-xl/stable-diffusion-xl
bash ./tools/measure.sh
popd

Performance Optimization with FP8 Flow

To optimize performance, we set heavy-performance ops to operate in FP8-143. In the conversion to FP8-143 we use various values of exponent bias which are determined using a calibration dataset. For each input processed, the UNET block is iteratively invoked 20 times. A more aggressive form is used for the first 18 steps; and a less aggressive one for the final 2 steps.

Environment Variables

All necessary environmental variables are enabled by default.

Supported Configurations

Validated on Intel Gaudi Software Version Framework Version(s) Mode
Gaudi 2 1.17.1 PyTorch 2.3.1 Inference