Text Summarization using LLAMA2-70b

Install CM following the installation page.

MLCommons-PythonNvidiaNeural Magic

MLPerf Reference Implementation in Python

Tip

• MLCommons reference implementations are only meant to provide a rules compliant reference implementation for the submitters and in most cases are not best performing. If you want to benchmark any system, it is advisable to use the vendor MLPerf implementation for that system like Nvidia, Intel etc.

LLAMA2-70B-99

datacenter

Datacenter category

In the datacenter category, llama2-70b-99 has Offline, Server scenarios and all the scenarios are mandatory for a closed division submission.

Pytorch

Pytorch framework

CPUCUDAROCm

CPU device

Please click here to see the minimum system requirements for running the benchmark

• Disk Space: 700GB

DockerNative

Docker Environment

Please refer to the installation page to install CM for running the automated benchmark commands.

# Docker Container Build and Performance Estimation for Offline Scenario

Tip

• Batch size could be adjusted using --batch_size=#, where # is the desired batch size. This option works only if the implementation in use is supporting the given batch size.

cm run script --tags=run-mlperf,inference,_find-performance,_full,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=test \
   --device=cpu  \
   --docker --quiet \
   --test_query_count=50

The above command should get you to an interactive shell inside the docker container and do a quick test run for the Offline scenario. Once inside the docker container please do the below commands to do the accuracy + performance runs for each scenario.

Please click here to see more options for the docker launch

• --docker_cm_repo=<Custom CM repo URL>: to use a custom fork of cm4mlops repository inside the docker image

• --docker_cache=no: to not use docker cache during the image build

• --docker_os=ubuntu: ubuntu and rhel are supported.
• --docker_os_version=20.04: [20.04, 22.04] are supported for Ubuntu and [8, 9] for RHEL

OfflineServerAll Scenarios

Offline

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=valid \
   --device=cpu \
   --quiet

Server

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Server\
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cpu \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

All Scenarios

cm run script --tags=run-mlperf,inference,_r4.1-dev,_all-scenarios \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cpu \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

Please click here to see more options for the RUN command

• Use --division=closed to do a closed division submission which includes compliance runs

• Use --rerun to do a rerun even when a valid run exists

Native Environment

Please refer to the installation page to install CM for running the automated benchmark commands.

# Setup a virtual environment for Python

cm run script --tags=install,python-venv --name=mlperf
export CM_SCRIPT_EXTRA_CMD="--adr.python.name=mlperf"

# Performance Estimation for Offline Scenario

cm run script --tags=run-mlperf,inference,_find-performance,_full,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=test \
   --device=cpu  \
   --quiet \
   --test_query_count=50

The above command should do a test run of Offline scenario and record the estimated offline_target_qps.

OfflineServerAll Scenarios

Offline

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=valid \
   --device=cpu \
   --quiet

Server

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Server\
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cpu \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

All Scenarios

cm run script --tags=run-mlperf,inference,_r4.1-dev,_all-scenarios \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cpu \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

Please click here to see more options for the RUN command

• Use --division=closed to do a closed division submission which includes compliance runs

• Use --rerun to do a rerun even when a valid run exists

CUDA device

Please click here to see the minimum system requirements for running the benchmark

• Device Memory: 8x80GB

• Disk Space: 700GB

DockerNative

Docker Environment

Please refer to the installation page to install CM for running the automated benchmark commands.

# Docker Container Build and Performance Estimation for Offline Scenario

Tip

• Batch size could be adjusted using --batch_size=#, where # is the desired batch size. This option works only if the implementation in use is supporting the given batch size.

cm run script --tags=run-mlperf,inference,_find-performance,_full,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=test \
   --device=cuda  \
   --docker --quiet \
   --test_query_count=50

The above command should get you to an interactive shell inside the docker container and do a quick test run for the Offline scenario. Once inside the docker container please do the below commands to do the accuracy + performance runs for each scenario.

Please click here to see more options for the docker launch

• --docker_cm_repo=<Custom CM repo URL>: to use a custom fork of cm4mlops repository inside the docker image

• --docker_cache=no: to not use docker cache during the image build

OfflineServerAll Scenarios

Offline

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=valid \
   --device=cuda \
   --quiet

Server

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Server\
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

All Scenarios

cm run script --tags=run-mlperf,inference,_r4.1-dev,_all-scenarios \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

Please click here to see more options for the RUN command

• Use --division=closed to do a closed division submission which includes compliance runs

• Use --rerun to do a rerun even when a valid run exists

Native Environment

Please refer to the installation page to install CM for running the automated benchmark commands.

# Setup a virtual environment for Python

cm run script --tags=install,python-venv --name=mlperf
export CM_SCRIPT_EXTRA_CMD="--adr.python.name=mlperf"

# Performance Estimation for Offline Scenario

cm run script --tags=run-mlperf,inference,_find-performance,_full,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=test \
   --device=cuda  \
   --quiet \
   --test_query_count=50

The above command should do a test run of Offline scenario and record the estimated offline_target_qps.

OfflineServerAll Scenarios

Offline

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=valid \
   --device=cuda \
   --quiet

Server

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Server\
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

All Scenarios

cm run script --tags=run-mlperf,inference,_r4.1-dev,_all-scenarios \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

Please click here to see more options for the RUN command

• Use --division=closed to do a closed division submission which includes compliance runs

• Use --rerun to do a rerun even when a valid run exists

ROCm device

Please click here to see the minimum system requirements for running the benchmark

• Disk Space: 700GB

Native

Native Environment

Please refer to the installation page to install CM for running the automated benchmark commands.

# Setup a virtual environment for Python

cm run script --tags=install,python-venv --name=mlperf
export CM_SCRIPT_EXTRA_CMD="--adr.python.name=mlperf"

# Performance Estimation for Offline Scenario

cm run script --tags=run-mlperf,inference,_find-performance,_full,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=test \
   --device=rocm  \
   --quiet \
   --test_query_count=50

The above command should do a test run of Offline scenario and record the estimated offline_target_qps.

OfflineServerAll Scenarios

Offline

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=valid \
   --device=rocm \
   --quiet

Server

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Server\
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=rocm \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

All Scenarios

cm run script --tags=run-mlperf,inference,_r4.1-dev,_all-scenarios \
   --model=llama2-70b-99 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=rocm \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

Please click here to see more options for the RUN command

• Use --division=closed to do a closed division submission which includes compliance runs

• Use --rerun to do a rerun even when a valid run exists

LLAMA2-70B-99.9

datacenter

Datacenter category

In the datacenter category, llama2-70b-99.9 has Offline, Server scenarios and all the scenarios are mandatory for a closed division submission.

Pytorch

Pytorch framework

CPUCUDAROCm

CPU device

Please click here to see the minimum system requirements for running the benchmark

• Disk Space: 700GB

DockerNative

Docker Environment

Please refer to the installation page to install CM for running the automated benchmark commands.

You can reuse the same environment as described for llama2-70b-99.

Performance Estimation for Offline Scenario

cm run script --tags=run-mlperf,inference,_find-performance,_full,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=test \
   --device=cpu  \
   --quiet \
   --test_query_count=50

The above command should do a test run of Offline scenario and record the estimated offline_target_qps.

OfflineServerAll Scenarios

Offline

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=valid \
   --device=cpu \
   --quiet

Server

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Server\
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cpu \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

All Scenarios

cm run script --tags=run-mlperf,inference,_r4.1-dev,_all-scenarios \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cpu \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

Please click here to see more options for the RUN command

• Use --division=closed to do a closed division submission which includes compliance runs

• Use --rerun to do a rerun even when a valid run exists

Native Environment

Please refer to the installation page to install CM for running the automated benchmark commands.

You can reuse the same environment as described for llama2-70b-99.

Performance Estimation for Offline Scenario

cm run script --tags=run-mlperf,inference,_find-performance,_full,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=test \
   --device=cpu  \
   --quiet \
   --test_query_count=50

The above command should do a test run of Offline scenario and record the estimated offline_target_qps.

OfflineServerAll Scenarios

Offline

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=valid \
   --device=cpu \
   --quiet

Server

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Server\
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cpu \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

All Scenarios

cm run script --tags=run-mlperf,inference,_r4.1-dev,_all-scenarios \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cpu \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

Please click here to see more options for the RUN command

• Use --division=closed to do a closed division submission which includes compliance runs

• Use --rerun to do a rerun even when a valid run exists

CUDA device

Please click here to see the minimum system requirements for running the benchmark

• Device Memory: 8x80GB

• Disk Space: 700GB

DockerNative

Docker Environment

Please refer to the installation page to install CM for running the automated benchmark commands.

You can reuse the same environment as described for llama2-70b-99.

Performance Estimation for Offline Scenario

cm run script --tags=run-mlperf,inference,_find-performance,_full,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=test \
   --device=cuda  \
   --quiet \
   --test_query_count=50

The above command should do a test run of Offline scenario and record the estimated offline_target_qps.

OfflineServerAll Scenarios

Offline

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=valid \
   --device=cuda \
   --quiet

Server

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Server\
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

All Scenarios

cm run script --tags=run-mlperf,inference,_r4.1-dev,_all-scenarios \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

Please click here to see more options for the RUN command

• Use --division=closed to do a closed division submission which includes compliance runs

• Use --rerun to do a rerun even when a valid run exists

Native Environment

Please refer to the installation page to install CM for running the automated benchmark commands.

You can reuse the same environment as described for llama2-70b-99.

Performance Estimation for Offline Scenario

cm run script --tags=run-mlperf,inference,_find-performance,_full,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=test \
   --device=cuda  \
   --quiet \
   --test_query_count=50

The above command should do a test run of Offline scenario and record the estimated offline_target_qps.

OfflineServerAll Scenarios

Offline

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=valid \
   --device=cuda \
   --quiet

Server

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Server\
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

All Scenarios

cm run script --tags=run-mlperf,inference,_r4.1-dev,_all-scenarios \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

Please click here to see more options for the RUN command

• Use --division=closed to do a closed division submission which includes compliance runs

• Use --rerun to do a rerun even when a valid run exists

ROCm device

Please click here to see the minimum system requirements for running the benchmark

• Disk Space: 700GB

Native

Native Environment

Please refer to the installation page to install CM for running the automated benchmark commands.

You can reuse the same environment as described for llama2-70b-99.

Performance Estimation for Offline Scenario

cm run script --tags=run-mlperf,inference,_find-performance,_full,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=test \
   --device=rocm  \
   --quiet \
   --test_query_count=50

The above command should do a test run of Offline scenario and record the estimated offline_target_qps.

OfflineServerAll Scenarios

Offline

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=valid \
   --device=rocm \
   --quiet

Server

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Server\
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=rocm \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

All Scenarios

cm run script --tags=run-mlperf,inference,_r4.1-dev,_all-scenarios \
   --model=llama2-70b-99.9 \
   --implementation=reference \
   --framework=pytorch \
   --category=datacenter \
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=rocm \
   --quiet

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

Please click here to see more options for the RUN command

• Use --division=closed to do a closed division submission which includes compliance runs

• Use --rerun to do a rerun even when a valid run exists

• If you want to download the official MLPerf model and dataset for llama2-70b-99.9 you can follow this README.

Nvidia MLPerf Implementation

LLAMA2-70B-99

datacenter

Datacenter category

In the datacenter category, llama2-70b-99 has Offline, Server scenarios and all the scenarios are mandatory for a closed division submission.

TensorRT

TensorRT framework

CUDA

CUDA device

Please click here to see the minimum system requirements for running the benchmark

• Device Memory: 80GB

• Disk Space: 700GB

Docker

Docker Environment

Please refer to the installation page to install CM for running the automated benchmark commands.

# Docker Container Build and Performance Estimation for Offline Scenario

Tip

• Batch size could be adjusted using --batch_size=#, where # is the desired batch size. This option works only if the implementation in use is supporting the given batch size.

• Default batch size is assigned based on GPU memory or the specified GPU. Please click more option for docker launch or run command to see how to specify the GPU name.

• When run with --all_models=yes, all the benchmark models of NVIDIA implementation can be executed within the same container.

• The dataset for NVIDIA's implementation of Llama2 is not publicly available. The user must fill this form and be verified as a MLCommons member to access the dataset.

• PATH_TO_PICKE_FILE should be replaced with path to the downloaded pickle file.

cm run script --tags=run-mlperf,inference,_find-performance,_full,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=nvidia \
   --framework=tensorrt \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=test \
   --device=cuda  \
   --docker --quiet \
   --test_query_count=50 \
   --tp_size=2 \
   --nvidia_llama2_dataset_file_path=<PATH_TO_PICKE_FILE>

The above command should get you to an interactive shell inside the docker container and do a quick test run for the Offline scenario. Once inside the docker container please do the below commands to do the accuracy + performance runs for each scenario.

Please click here to see more options for the docker launch

• --docker_cm_repo=<Custom CM repo URL>: to use a custom fork of cm4mlops repository inside the docker image

• --docker_cache=no: to not use docker cache during the image build

• --gpu_name=<Name of the GPU> : The GPUs with supported configs in CM are orin, rtx_4090, rtx_a6000, rtx_6000_ada, l4, t4and a100. For other GPUs, default configuration as per the GPU memory will be used.

OfflineServerAll Scenarios

Offline

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=nvidia \
   --framework=tensorrt \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=valid \
   --device=cuda \
   --quiet \
   --tp_size=<TP_SIZE> \
   --nvidia_llama2_dataset_file_path=<PATH_TO_PICKE_FILE>

Server

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=nvidia \
   --framework=tensorrt \
   --category=datacenter \
   --scenario=Server\
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet \
   --tp_size=<TP_SIZE> \
   --nvidia_llama2_dataset_file_path=<PATH_TO_PICKE_FILE>

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

All Scenarios

cm run script --tags=run-mlperf,inference,_r4.1-dev,_all-scenarios \
   --model=llama2-70b-99 \
   --implementation=nvidia \
   --framework=tensorrt \
   --category=datacenter \
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet \
   --tp_size=<TP_SIZE> \
   --nvidia_llama2_dataset_file_path=<PATH_TO_PICKE_FILE>

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

Please click here to see more options for the RUN command

• Use --division=closed to do a closed division submission which includes compliance runs

• Use --rerun to do a rerun even when a valid run exists

• --gpu_name=<Name of the GPU> : The GPUs with supported configs in CM are orin, rtx_4090, rtx_a6000, rtx_6000_ada, l4, t4and a100. For other GPUs, default configuration as per the GPU memory will be used.

LLAMA2-70B-99.9

datacenter

Datacenter category

In the datacenter category, llama2-70b-99.9 has Offline, Server scenarios and all the scenarios are mandatory for a closed division submission.

TensorRT

TensorRT framework

CUDA

CUDA device

Please click here to see the minimum system requirements for running the benchmark

• Device Memory: 80GB

• Disk Space: 700GB

Docker

Docker Environment

Please refer to the installation page to install CM for running the automated benchmark commands.

You can reuse the same environment as described for llama2-70b-99.

Performance Estimation for Offline Scenario

cm run script --tags=run-mlperf,inference,_find-performance,_full,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=nvidia \
   --framework=tensorrt \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=test \
   --device=cuda  \
   --quiet \
   --test_query_count=50 \
   --tp_size=2 \
   --nvidia_llama2_dataset_file_path=<PATH_TO_PICKE_FILE>

The above command should do a test run of Offline scenario and record the estimated offline_target_qps.

OfflineServerAll Scenarios

Offline

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=nvidia \
   --framework=tensorrt \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=valid \
   --device=cuda \
   --quiet \
   --tp_size=<TP_SIZE> \
   --nvidia_llama2_dataset_file_path=<PATH_TO_PICKE_FILE>

Server

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=nvidia \
   --framework=tensorrt \
   --category=datacenter \
   --scenario=Server\
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet \
   --tp_size=<TP_SIZE> \
   --nvidia_llama2_dataset_file_path=<PATH_TO_PICKE_FILE>

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

All Scenarios

cm run script --tags=run-mlperf,inference,_r4.1-dev,_all-scenarios \
   --model=llama2-70b-99.9 \
   --implementation=nvidia \
   --framework=tensorrt \
   --category=datacenter \
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet \
   --tp_size=<TP_SIZE> \
   --nvidia_llama2_dataset_file_path=<PATH_TO_PICKE_FILE>

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

Please click here to see more options for the RUN command

• Use --division=closed to do a closed division submission which includes compliance runs

• Use --rerun to do a rerun even when a valid run exists

• --gpu_name=<Name of the GPU> : The GPUs with supported configs in CM are orin, rtx_4090, rtx_a6000, rtx_6000_ada, l4, t4and a100. For other GPUs, default configuration as per the GPU memory will be used.

Neural Magic MLPerf Implementation

LLAMA2-70B-99

datacenter

Datacenter category

In the datacenter category, llama2-70b-99 has Offline, Server scenarios and all the scenarios are mandatory for a closed division submission.

pytorch

pytorch framework

CUDA

CUDA device

Please click here to see the minimum system requirements for running the benchmark

• Disk Space: 700GB

DockerNative

Docker Environment

Please refer to the installation page to install CM for running the automated benchmark commands.

# Run the Inference Server

cm run script --tags=run,vllm-server \
 --model=nm-testing/Llama-2-70b-chat-hf-FP8 \
 --vllm_model_name=nm-testing/Llama-2-70b-chat-hf-FP8 \
 --quiet

Tip

• Host and Port number of the server can be configured through --host and --port options. Otherwise, server will run on the default host localhost and port 8000.

# Docker Container Build and Performance Estimation for Offline Scenario

Tip

• Batch size could be adjusted using --batch_size=#, where # is the desired batch size. This option works only if the implementation in use is supporting the given batch size.

cm run script --tags=run-mlperf,inference,_find-performance,_full,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=neuralmagic \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=test \
   --device=cuda  \
   --docker --quiet \
   --test_query_count=50 \
   --api_server=http://localhost:8000 \
   --vllm_model_name=nm-testing/Llama-2-70b-chat-hf-FP8 \
   --adr.mlperf-implementation.tags=_repo.https://github.com/neuralmagic/inference,_branch.vllm

The above command should get you to an interactive shell inside the docker container and do a quick test run for the Offline scenario. Once inside the docker container please do the below commands to do the accuracy + performance runs for each scenario.

Please click here to see more options for the docker launch

• --docker_cm_repo=<Custom CM repo URL>: to use a custom fork of cm4mlops repository inside the docker image

• --docker_cache=no: to not use docker cache during the image build

OfflineServerAll Scenarios

Offline

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=neuralmagic \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=valid \
   --device=cuda \
   --quiet \
   --api_server=http://localhost:8000 \
   --vllm_model_name=nm-testing/Llama-2-70b-chat-hf-FP8 \
   --adr.mlperf-implementation.tags=_repo.https://github.com/neuralmagic/inference,_branch.vllm

Server

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=neuralmagic \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Server\
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet \
   --api_server=http://localhost:8000 \
   --vllm_model_name=nm-testing/Llama-2-70b-chat-hf-FP8 \
   --adr.mlperf-implementation.tags=_repo.https://github.com/neuralmagic/inference,_branch.vllm

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

All Scenarios

cm run script --tags=run-mlperf,inference,_r4.1-dev,_all-scenarios \
   --model=llama2-70b-99 \
   --implementation=neuralmagic \
   --framework=pytorch \
   --category=datacenter \
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet \
   --api_server=http://localhost:8000 \
   --vllm_model_name=nm-testing/Llama-2-70b-chat-hf-FP8 \
   --adr.mlperf-implementation.tags=_repo.https://github.com/neuralmagic/inference,_branch.vllm

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

Please click here to see more options for the RUN command

• Use --division=closed to do a closed division submission which includes compliance runs

• Use --rerun to do a rerun even when a valid run exists

Native Environment

Please refer to the installation page to install CM for running the automated benchmark commands.

# Run the Inference Server

cm run script --tags=run,vllm-server \
 --model=nm-testing/Llama-2-70b-chat-hf-FP8 \
 --vllm_model_name=nm-testing/Llama-2-70b-chat-hf-FP8 \
 --quiet

Tip

• Host and Port number of the server can be configured through --host and --port options. Otherwise, server will run on the default host localhost and port 8000.

# Setup a virtual environment for Python

cm run script --tags=install,python-venv --name=mlperf
export CM_SCRIPT_EXTRA_CMD="--adr.python.name=mlperf"

# Performance Estimation for Offline Scenario

cm run script --tags=run-mlperf,inference,_find-performance,_full,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=neuralmagic \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=test \
   --device=cuda  \
   --quiet \
   --test_query_count=50 \
   --api_server=http://localhost:8000 \
   --vllm_model_name=nm-testing/Llama-2-70b-chat-hf-FP8 \
   --adr.mlperf-implementation.tags=_repo.https://github.com/neuralmagic/inference,_branch.vllm

The above command should do a test run of Offline scenario and record the estimated offline_target_qps.

OfflineServerAll Scenarios

Offline

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=neuralmagic \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=valid \
   --device=cuda \
   --quiet \
   --api_server=http://localhost:8000 \
   --vllm_model_name=nm-testing/Llama-2-70b-chat-hf-FP8 \
   --adr.mlperf-implementation.tags=_repo.https://github.com/neuralmagic/inference,_branch.vllm

Server

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99 \
   --implementation=neuralmagic \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Server\
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet \
   --api_server=http://localhost:8000 \
   --vllm_model_name=nm-testing/Llama-2-70b-chat-hf-FP8 \
   --adr.mlperf-implementation.tags=_repo.https://github.com/neuralmagic/inference,_branch.vllm

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

All Scenarios

cm run script --tags=run-mlperf,inference,_r4.1-dev,_all-scenarios \
   --model=llama2-70b-99 \
   --implementation=neuralmagic \
   --framework=pytorch \
   --category=datacenter \
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet \
   --api_server=http://localhost:8000 \
   --vllm_model_name=nm-testing/Llama-2-70b-chat-hf-FP8 \
   --adr.mlperf-implementation.tags=_repo.https://github.com/neuralmagic/inference,_branch.vllm

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

Please click here to see more options for the RUN command

• Use --division=closed to do a closed division submission which includes compliance runs

• Use --rerun to do a rerun even when a valid run exists

LLAMA2-70B-99.9

datacenter

Datacenter category

In the datacenter category, llama2-70b-99.9 has Offline, Server scenarios and all the scenarios are mandatory for a closed division submission.

pytorch

pytorch framework

CUDA

CUDA device

Please click here to see the minimum system requirements for running the benchmark

• Disk Space: 700GB

DockerNative

Docker Environment

Please refer to the installation page to install CM for running the automated benchmark commands.

You can reuse the same environment as described for llama2-70b-99.

Performance Estimation for Offline Scenario

cm run script --tags=run-mlperf,inference,_find-performance,_full,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=neuralmagic \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=test \
   --device=cuda  \
   --quiet \
   --test_query_count=50 \
   --api_server=http://localhost:8000 \
   --vllm_model_name=nm-testing/Llama-2-70b-chat-hf-FP8 \
   --adr.mlperf-implementation.tags=_repo.https://github.com/neuralmagic/inference,_branch.vllm

The above command should do a test run of Offline scenario and record the estimated offline_target_qps.

OfflineServerAll Scenarios

Offline

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=neuralmagic \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=valid \
   --device=cuda \
   --quiet \
   --api_server=http://localhost:8000 \
   --vllm_model_name=nm-testing/Llama-2-70b-chat-hf-FP8 \
   --adr.mlperf-implementation.tags=_repo.https://github.com/neuralmagic/inference,_branch.vllm

Server

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=neuralmagic \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Server\
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet \
   --api_server=http://localhost:8000 \
   --vllm_model_name=nm-testing/Llama-2-70b-chat-hf-FP8 \
   --adr.mlperf-implementation.tags=_repo.https://github.com/neuralmagic/inference,_branch.vllm

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

All Scenarios

cm run script --tags=run-mlperf,inference,_r4.1-dev,_all-scenarios \
   --model=llama2-70b-99.9 \
   --implementation=neuralmagic \
   --framework=pytorch \
   --category=datacenter \
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet \
   --api_server=http://localhost:8000 \
   --vllm_model_name=nm-testing/Llama-2-70b-chat-hf-FP8 \
   --adr.mlperf-implementation.tags=_repo.https://github.com/neuralmagic/inference,_branch.vllm

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

Please click here to see more options for the RUN command

• Use --division=closed to do a closed division submission which includes compliance runs

• Use --rerun to do a rerun even when a valid run exists

Native Environment

Please refer to the installation page to install CM for running the automated benchmark commands.

You can reuse the same environment as described for llama2-70b-99.

Performance Estimation for Offline Scenario

cm run script --tags=run-mlperf,inference,_find-performance,_full,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=neuralmagic \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=test \
   --device=cuda  \
   --quiet \
   --test_query_count=50 \
   --api_server=http://localhost:8000 \
   --vllm_model_name=nm-testing/Llama-2-70b-chat-hf-FP8 \
   --adr.mlperf-implementation.tags=_repo.https://github.com/neuralmagic/inference,_branch.vllm

The above command should do a test run of Offline scenario and record the estimated offline_target_qps.

OfflineServerAll Scenarios

Offline

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=neuralmagic \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Offline \
   --execution_mode=valid \
   --device=cuda \
   --quiet \
   --api_server=http://localhost:8000 \
   --vllm_model_name=nm-testing/Llama-2-70b-chat-hf-FP8 \
   --adr.mlperf-implementation.tags=_repo.https://github.com/neuralmagic/inference,_branch.vllm

Server

cm run script --tags=run-mlperf,inference,_r4.1-dev \
   --model=llama2-70b-99.9 \
   --implementation=neuralmagic \
   --framework=pytorch \
   --category=datacenter \
   --scenario=Server\
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet \
   --api_server=http://localhost:8000 \
   --vllm_model_name=nm-testing/Llama-2-70b-chat-hf-FP8 \
   --adr.mlperf-implementation.tags=_repo.https://github.com/neuralmagic/inference,_branch.vllm

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

All Scenarios

cm run script --tags=run-mlperf,inference,_r4.1-dev,_all-scenarios \
   --model=llama2-70b-99.9 \
   --implementation=neuralmagic \
   --framework=pytorch \
   --category=datacenter \
   --server_target_qps=<SERVER_TARGET_QPS> \
   --execution_mode=valid \
   --device=cuda \
   --quiet \
   --api_server=http://localhost:8000 \
   --vllm_model_name=nm-testing/Llama-2-70b-chat-hf-FP8 \
   --adr.mlperf-implementation.tags=_repo.https://github.com/neuralmagic/inference,_branch.vllm

Tip

• <SERVER_TARGET_QPS> must be determined manually. It is usually around 80% of the Offline QPS, but on some systems, it can drop below 50%. If a higher value is specified, the latency constraint will not be met, and the run will be considered invalid.

Please click here to see more options for the RUN command

• Use --division=closed to do a closed division submission which includes compliance runs

• Use --rerun to do a rerun even when a valid run exists