Habana MLPerf™ inference submission

This directory provides instructions to reproduce Habana'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.

Habana MLPerf™ inference submission

Setup

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

Prepare MLPerf Directory

Perform the following:

Follow the instructions provided in the Gaudi Installation Guide to set up the environment including the $PYTHON environment variable. The guide will walk you through the process of setting up your system to run the benchmarks on Gaudi.

Clone Model-References repository and switch to the branch that matches your SynapseAI version. You can run the hl-smi utility to determine the SynapseAI version.

export MLPERF_ROOT=/path/to/mlperf/root
cd $MLPERF_ROOT
git clone -b [SynapseAI version] https://github.com/HabanaAI/Model-References
export MLPERF_DIR=$MLPERF_ROOT/Model-References/MLPERF3.1/Inference

Build and Deploy HabanaLabs Container

To build MLPerf inference 3.1 container, perform the following:

Set the environment variables for the docker command.
- To find a docker image, go to gaudi-docker.
- Open gaudi-docker directory, and select the folder that matches the SynapseAI version (determined by running hl-smi).
- Navigate to subdirectories, choose system and framework version.
- Choose the docker build version. Most often 'latest' will be used.
- Navigate to "Docker Info" tab and note "Title" string.
- Set DOCKER_IMAGE to "Title" string with vault.habana.ai/gaudi-docker/ prefix. See the examples below.
  - Example on PyTorch Container:
```
# NOTE: The below is only an example value. Replace [SynapseAI version] and [PT version] to match your setup and Supported Configuration.
export DOCKER_IMAGE=vault.habana.ai/gaudi-docker/[SynapseAI version]/ubuntu20.04/habanalabs/pytorch-installer-[PT Version]:latest
```
Create mlperf-habana container by running the following command.

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 $MLPERF_DIR:/root/Habana/            \
           --cap-add=sys_nice --cap-add=SYS_PTRACE \
           --user root --workdir=/root --net=host  \
           --ulimit memlock=-1:-1 ${DOCKER_IMAGE}

Start the docker.

docker exec -it mlperf-habana bash

Download Checkpoint

mkdir -p /mnt/weka/data/pytorch/
pushd /mnt/weka/data/pytorch/
wget https://cloud.mlcommons.org/index.php/s/QAZ2oM94MkFtbQx/download  --output-document checkpoint.zip
unzip -q checkpoint.zip && rm checkpoint.zip
popd

Download Dataset

pushd /root/Habana/code/gptj-99.9/gpt-j
python download_cnndm.py
cp data/cnn_eval.json /mnt/weka/data/pytorch/gpt-j/cnn_eval.json
popd

Reproduce Results

99 and 99.9 Accuracy

The same script was submitted for both 99 and 99.9 benchmarks - no additional improvements were made for low accuracy (99), and 99.9 results were used for 99 as well.

Get Started

Install the requirements and build the latest loadgen.

cd /root/Habana/code
source functions.sh
build_mlperf_inference

Generate Results

To generate full submission results, run the following command:

build_mlperf_inference --output-dir <path_to_output_dir> --submission gptj-99.9-fp8

The command produces results from accuracy and performance runs for both Offline and Server scenarios. Logs can be found under /output_dir/logs/model/, e.g. /results/logs/gptj-99.9-fp8/

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

build_mlperf_inference --output-dir <path_to_output_dir> --submission gptj-99.9-fp8_Offline

build_mlperf_inference --output-dir <path_to_output_dir> --submission gptj-99.9-fp8_Server

Logs can be found under /output_dir/logs/model/scenario/, e.g. /results/logs/gptj-99.9-fp8/Offline/

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

build_mlperf_inference --output-dir <path_to_output_dir> --submission gptj-99.9-fp8_Server --mode acc

build_mlperf_inference --output-dir <path_to_output_dir> --submission gptj-99.9-fp8_Offline --mode perf

Logs can be found under /output_dir/logs/model/scenario/mode/, e.g. /results/logs/gptj-99.9-fp8/Offline/accuracy/

Performance Optimization with FP8 Flow

To optimize performance, we set heavy-performance ops to operate in fp8-143.

All fp8 ops are working with a fixed fp8 exponent bias = 7 and no scaling is required.

Environment Variables

The following outlines custom ENV variables used in the GPT-J submission script:

| Enviroment Variable | Effect | |------------------------------------------------------------------------- |------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | PT_USE_FP8_143=1 | Sets PT backend fp8 flavor to fp8_143 | | UPDATE_MME_OUTPUT_PRECISION_FILTER="v_proj,matmul_av" | Allows the specified MME layer to output fp8 for performance optimization. | | SCALES_FILE_PATH=quantization/measurements/per_tensor_scales_gpt_j.json | Loads per-tensor scales required for fp8 quantization. If not provided, no scaling is applied. | | ENABLE_EXPERIMENTAL_FLAGS=true | Enables the above flags |

Supported Configurations

Validated on	SynapseAI Version	Framework Version(s)	Mode
Gaudi2	1.14.0	PyTorch 2.1.1	Inference

Changelog

1.13.0

Published MLPerf™ inference 3.1 GPT-J script