git init

graph/R-GAT/tools/download_igbh_full.sh

+#!/bin/bash
+# This script was taken from: https://github.com/IllinoisGraphBenchmark/IGB-Datasets/blob/main/igb/download_igbh600m.sh
+# Copy of licence for this ONLY this script
+# MIT License
+
+# Copyright (c) 2022 IMPACT Research Group and IllinoisGraphBenchmark
+
+# This license is applicable for all software codebases provided in the repository.
+
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+echo "IGBH600M download starting"
+mkdir -p $1/full/processed
+cd $1/full/processed
+
+# paper
+mkdir paper
+cd paper
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/paper/node_feat.npy
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/paper/node_label_19.npy
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/paper/node_label_2K.npy
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/paper/paper_id_index_mapping.npy
+cd ..
+
+# paper__cites__paper
+mkdir paper__cites__paper
+cd paper__cites__paper
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/paper__cites__paper/edge_index.npy
+cd ..
+
+# author
+mkdir author
+cd author
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/author/author_id_index_mapping.npy
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/author/node_feat.npy
+cd ..
+
+# conference
+mkdir conference
+cd conference
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/conference/conference_id_index_mapping.npy
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/conference/node_feat.npy
+cd ..
+
+# institute
+mkdir institute
+cd institute
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/institute/institute_id_index_mapping.npy
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/institute/node_feat.npy
+cd ..
+
+# journal
+mkdir journal
+cd journal
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/journal/journal_id_index_mapping.npy
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/journal/node_feat.npy
+cd ..
+
+# fos
+mkdir fos
+cd fos
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/fos/fos_id_index_mapping.npy
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/fos/node_feat.npy
+cd ..
+
+# author__affiliated_to__institute
+mkdir author__affiliated_to__institute
+cd author__affiliated_to__institute
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/author__affiliated_to__institute/edge_index.npy
+cd ..
+
+# paper__published__journal
+mkdir paper__published__journal
+cd paper__published__journal
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/paper__published__journal/edge_index.npy
+cd ..
+
+# paper__topic__fos
+mkdir paper__topic__fos
+cd paper__topic__fos
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/paper__topic__fos/edge_index.npy
+cd ..
+
+# paper__venue__conference
+mkdir paper__venue__conference
+cd paper__venue__conference
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/paper__venue__conference/edge_index.npy
+cd ..
+
+# paper__written_by__author
+mkdir paper__written_by__author
+cd paper__written_by__author
+wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/paper__written_by__author/edge_index.npy
+cd ..
+
+echo "IGBH-IGBH download complete"
\ No newline at end of file

graph/R-GAT/tools/download_igbh_test.py

+import argparse
+import os
+from igb import download
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--target-path",
+        type=str,
+        default="igbh/",
+        help="Download path for the dataset",
+    )
+    parser.add_argument(
+        "--dataset-size",
+        type=str,
+        default="tiny",
+        choices=["tiny", "small", "medium"],
+        help="Size of the dataset, only full for official submissions",
+    )
+    parser.add_argument(
+        "--dataset-type",
+        type=str,
+        default="heterogeneous",
+        choices=["homogeneous", "heterogeneous"],
+    )
+
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    args = get_args()
+    os.makedirs(args.target_path, exist_ok=True)
+    download.download_dataset(
+        path=args.target_path,
+        dataset_size=args.dataset_size,
+        dataset_type=args.dataset_type,
+    )

graph/R-GAT/tools/split_seeds.py

+# Script taken from
+# https://github.com/mlcommons/training/blob/master/graph_neural_network/split_seeds.py
+
+import argparse
+import os.path as osp
+import torch
+
+
+class SeedSplitter(object):
+    def __init__(
+        self,
+        path,
+        dataset_size="tiny",
+        use_label_2K=True,
+        random_seed=42,
+        validation_frac=0.01,
+        calibration=False
+    ):
+        self.path = path
+        self.dataset_size = dataset_size
+        self.use_label_2K = use_label_2K
+        self.random_seed = random_seed
+        self.validation_frac = validation_frac
+        self.calibration = calibration
+        self.paper_nodes_num = {
+            "tiny": 100000,
+            "small": 1000000,
+            "medium": 10000000,
+            "large": 100000000,
+            "full": 269346174,
+        }
+        self.process()
+
+    def process(self):
+        torch.manual_seed(self.random_seed)
+        n_labeled_idx = self.paper_nodes_num[self.dataset_size]
+        if self.dataset_size == "full":
+            if self.use_label_2K:
+                n_labeled_idx = 157675969
+            else:
+                n_labeled_idx = 227130858
+
+        shuffled_index = torch.randperm(n_labeled_idx)
+        n_train = int(n_labeled_idx * 0.6)
+        n_val = int(n_labeled_idx * self.validation_frac)
+
+        train_idx = shuffled_index[:n_train]
+        val_idx = shuffled_index[n_train: n_train + n_val]
+
+        path = osp.join(self.path, self.dataset_size, "processed")
+        torch.save(train_idx, osp.join(path, "train_idx.pt"))
+        torch.save(val_idx, osp.join(path, "val_idx.pt"))
+
+        if self.calibration and self.dataset_size == "full":
+            n_calibration = 5000
+            calibration_idx = shuffled_index[:n_calibration].numpy().tolist()
+            with open(osp.join(path, "calibration.txt"), "w+") as f:
+                f.writelines([f"{idx}\n" for idx in calibration_idx])
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    root = osp.join(
+        osp.dirname(
+            osp.dirname(
+                osp.dirname(
+                    osp.realpath(__file__)))), "data", "igbh"
+    )
+    parser.add_argument(
+        "--path", type=str, default=root, help="path containing the datasets"
+    )
+    parser.add_argument(
+        "--dataset_size",
+        type=str,
+        default="full",
+        choices=["tiny", "small", "medium", "large", "full"],
+        help="size of the datasets",
+    )
+    parser.add_argument("--random_seed", type=int, default="42")
+    parser.add_argument(
+        "--num_classes",
+        type=int,
+        default=2983,
+        choices=[19, 2983],
+        help="number of classes",
+    )
+    parser.add_argument(
+        "--validation_frac",
+        type=float,
+        default=0.005,
+        help="Fraction of labeled vertices to be used for validation.",
+    )
+    parser.add_argument(
+        "--calibration",
+        action="store_true",
+        help="Save calibration dataset",
+    )
+
+    args = parser.parse_args()
+    splitter = SeedSplitter(
+        path=args.path,
+        dataset_size=args.dataset_size,
+        use_label_2K=(args.num_classes == 2983),
+        random_seed=args.random_seed,
+        validation_frac=args.validation_frac,
+        calibration=args.calibration,
+    )

graph/R-GAT/user.conf

+# The format of this config file is 'key = value'.
+# The key has the format 'model.scenario.key'. Value is mostly int64_t.
+# Model maybe '*' as wildcard. In that case the value applies to all models.
+# All times are in milli seconds

language/bert/.gitignore

+build/
+eval_features.pickle
+onnxruntime_profile__*.json

language/bert/Dockerfile

+# Copyright (c) 2019, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+ARG BASE_IMAGE
+FROM ${BASE_IMAGE}
+
+# Install LoadGen
+# Cloning the LoadGen so that we have clean git repo from within the docker container.
+RUN cd /tmp \
+ && git clone https://github.com/mlperf/inference.git \
+ && cd inference \
+ && git submodule update --init third_party/pybind \
+ && cd loadgen \
+ && python3 setup.py install \
+ && cd /tmp \
+ && rm -rf inference
+
+# Install dependencies
+RUN python3 -m pip install torch==1.4.0 onnx==1.6.0 transformers==2.4.0 \ 
+    onnxruntime==1.2.0 numpy==1.18.0 tokenization==1.0.7
+
+# Add user
+ARG GID
+ARG UID
+ARG GROUP
+ARG USER
+RUN echo root:root | chpasswd \
+ && groupadd -f -g ${GID} ${GROUP} \
+ && useradd -G sudo -g ${GID} -u ${UID} -m ${USER} \
+ && echo ${USER}:${USER} | chpasswd \
+ && echo -e "\nexport PS1=\"(mlperf) \\u@\\h:\\w\\$ \"" | tee -a /home/${USER}/.bashrc \
+ && echo -e "\n%sudo ALL=(ALL:ALL) NOPASSWD:ALL\n" | tee -a /etc/sudoers

language/bert/Makefile

+# Copyright (c) 2019, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+SHELL := /bin/bash
+
+MAKEFILE_NAME := $(lastword $(MAKEFILE_LIST))
+UNAME := $(shell whoami)
+UID := $(shell id -u `whoami`)
+GROUPNAME := $(shell id -gn `whoami`)
+GROUPID := $(shell id -g `whoami`)
+
+HOST_VOL ?= ${PWD}
+CONTAINER_VOL ?= /workspace
+
+BUILD_DIR := build
+DATA_DIR := $(BUILD_DIR)/data
+BERT_DIR := $(DATA_DIR)/bert_tf_v1_1_large_fp32_384_v2
+RESULT_DIR := $(BUILD_DIR)/result
+MLPERF_CONF := $(BUILD_DIR)/mlperf.conf
+FEATURE_CACHE := eval_features.pickle
+
+# Handle different nvidia-docker version
+ifneq ($(wildcard /usr/bin/nvidia-docker),)
+	DOCKER_RUN_CMD := nvidia-docker run
+else
+	DOCKER_RUN_CMD := docker run --gpus=all
+endif
+
+.PHONY: setup
+setup:
+	@if [ ! -e $(BUILD_DIR) ]; then \
+		mkdir $(BUILD_DIR); \
+	fi
+	@if [ ! -e $(MLPERF_CONF) ]; then \
+		cp ../../mlperf.conf $(MLPERF_CONF); \
+	fi
+	@$(MAKE) -f $(MAKEFILE_NAME) init_submodule
+	@$(MAKE) -f $(MAKEFILE_NAME) download_data
+	@$(MAKE) -f $(MAKEFILE_NAME) download_model
+
+.PHONY: init_submodule
+init_submodule:
+	@git submodule update --init DeepLearningExamples
+
+.PHONY: download_data
+download_data:
+	@if [ ! -e $(DATA_DIR) ]; then \
+		mkdir $(DATA_DIR); \
+	fi
+	@if [ ! -e $(DATA_DIR)/dev-v1.1.json ]; then \
+		wget -O $(DATA_DIR)/dev-v1.1.json https://github.com/rajpurkar/SQuAD-explorer/blob/master/dataset/dev-v1.1.json?raw=true; \
+	fi
+	@if [ ! -e $(DATA_DIR)/evaluate-v1.1.py ]; then \
+		wget -O $(DATA_DIR)/evaluate-v1.1.py https://github.com/allenai/bi-att-flow/raw/master/squad/evaluate-v1.1.py; \
+	fi
+	@if [ ! -e $(BERT_DIR) ]; then \
+		mkdir $(BERT_DIR) ; \
+	fi
+	@if [ ! -e $(RESULT_DIR) ]; then \
+		mkdir $(RESULT_DIR); \
+	fi
+
+.PHONY: download_model
+download_model:
+	@$(MAKE) -f $(MAKEFILE_NAME) download_tf_model
+	@$(MAKE) -f $(MAKEFILE_NAME) download_pytorch_model
+	@$(MAKE) -f $(MAKEFILE_NAME) download_onnx_model
+
+.PHONY: download_tf_model
+download_tf_model:
+	@if [ ! -e $(BERT_DIR)/model.ckpt-5474.data-00000-of-00001 ]; then \
+		wget -O $(BERT_DIR)/model.ckpt-5474.data-00000-of-00001 https://zenodo.org/record/3733868/files/model.ckpt-5474.data-00000-of-00001?download=1; \
+	fi
+	@if [ ! -e $(BERT_DIR)/model.ckpt-5474.index ]; then \
+		wget -O $(BERT_DIR)/model.ckpt-5474.index https://zenodo.org/record/3733868/files/model.ckpt-5474.index?download=1; \
+	fi
+	@if [ ! -e $(BERT_DIR)/model.ckpt-5474.meta ]; then \
+		wget -O $(BERT_DIR)/model.ckpt-5474.meta https://zenodo.org/record/3733868/files/model.ckpt-5474.meta?download=1; \
+	fi
+	@if [ ! -e $(BERT_DIR)/vocab.txt ]; then \
+		wget -O $(BERT_DIR)/vocab.txt https://zenodo.org/record/3733868/files/vocab.txt?download=1; \
+	fi
+	@if [ ! -e $(BERT_DIR)/model.pb ]; then \
+		wget -O $(BERT_DIR)/model.pb https://zenodo.org/record/3939747/files/model.pb?download=1; \
+	fi
+
+.PHONY: download_pytorch_model
+download_pytorch_model:
+	@if [ ! -e $(BERT_DIR)/model.pytorch ]; then \
+		wget -O $(BERT_DIR)/model.pytorch https://zenodo.org/record/3733896/files/model.pytorch?download=1; \
+	fi
+	@if [ ! -e $(BERT_DIR)/vocab.txt ]; then \
+		wget -O $(BERT_DIR)/vocab.txt https://zenodo.org/record/3733896/files/vocab.txt?download=1; \
+	fi
+
+.PHONY: download_onnx_model
+download_onnx_model:
+	@if [ ! -e $(BERT_DIR)/model.onnx ]; then \
+		wget -O $(BERT_DIR)/model.onnx https://zenodo.org/record/3733910/files/model.onnx?download=1; \
+	fi
+	@if [ ! -e $(BERT_DIR)/bert_large_v1_1_fake_quant.onnx ]; then \
+		wget -O $(BERT_DIR)/bert_large_v1_1_fake_quant.onnx https://zenodo.org/record/3750364/files/bert_large_v1_1_fake_quant.onnx?download=1; \
+	fi
+	@if [ ! -e $(BERT_DIR)/vocab.txt ]; then \
+		wget -O $(BERT_DIR)/vocab.txt https://zenodo.org/record/3733910/files/vocab.txt?download=1; \
+	fi
+
+.PHONY: build_docker
+build_docker:
+	@docker pull nvcr.io/nvidia/tensorrtserver:19.08-py3
+	@cd DeepLearningExamples/TensorFlow/LanguageModeling/BERT && docker build . --rm -t mlperf-inference-bert
+	@docker build --build-arg GID=$(GROUPID) --build-arg UID=$(UID) --build-arg GROUP=$(GROUPNAME) --build-arg USER=$(UNAME) \
+		--build-arg BASE_IMAGE=mlperf-inference-bert -t mlperf-inference-bert - < Dockerfile
+
+.PHONY: launch_docker
+launch_docker:
+	@$(DOCKER_RUN_CMD) --rm -it -w /workspace -v $(HOST_VOL):$(CONTAINER_VOL) -v ${HOME}:/mnt/${HOME} \
+		--shm-size=1g --ulimit memlock=-1 --ulimit stack=67108864 \
+		-v /etc/timezone:/etc/timezone:ro -v /etc/localtime:/etc/localtime:ro \
+		--security-opt apparmor=unconfined --security-opt seccomp=unconfined \
+		--name mlperf-inference-bert-$(UNAME) -h mlperf-inference-bert-$(UNAME) --add-host mlperf-inference-bert-$(UNAME):127.0.0.1 \
+		--user $(UID):$(GROUPID) --net host --device /dev/fuse --cap-add SYS_ADMIN $(DOCKER_ARGS) mlperf-inference-bert
+
+.PHONY: run_tf_performance
+run_tf_performance:
+	@python3 run.py --backend=tf
+
+.PHONY: run_tf_accuracy
+run_tf_accuracy:
+	@python3 run.py --backend=tf --accuracy
+
+.PHONY: run_pytorch_performance
+run_pytorch_performance:
+	@python3 run.py --backend=pytorch
+
+.PHONY: run_pytorch_accuracy
+run_pytorch_accuracy:
+	@python3 run.py --backend=pytorch --accuracy
+
+.PHONY: run_onnxruntime_performance
+run_onnxruntime_performance:
+	@python3 run.py --backend=onnxruntime
+
+.PHONY: run_onnxruntime_accuracy
+run_onnxruntime_accuracy:
+	@python3 run.py --backend=onnxruntime --accuracy
+
+.PHONY: run_onnxruntime_quantized_performance
+run_onnxruntime_quantized_performance:
+	@python3 run.py --backend=onnxruntime --quantized
+
+.PHONY: run_onnxruntime_quantized_accuracy
+run_onnxruntime_quantized_accuracy:
+	@python3 run.py --backend=onnxruntime --quantized --accuracy
+
+.PHONY: evaluate
+evaluate:
+	@python3 $(DATA_DIR)/evaluate-v1.1.py.json $(DATA_DIR)/dev-v1.1.json $(RESULT_DIR)/predictions.json
+
+.PHONY: clean
+clean:
+	@rm -rf ${BUILD_DIR}
+	@rm -f  ${FEATURE_CACHE}
+	@rm -f  onnxruntime_profile__*.json

language/bert/README.md

+# MLPerf Inference Benchmarks for Natural Language Processing
+
+This is the reference implementation for MLPerf Inference benchmarks for Natural Language Processing.
+
+The chosen model is BERT-Large performing SQuAD v1.1 question answering task.
+
+Please see the [new docs site](https://docs.mlcommons.org/inference/benchmarks/language/bert) for an automated way to run this benchmark across different available implementations and do an end-to-end submission with or without docker.
+
+## Prerequisites
+
+- [nvidia-docker](https://github.com/NVIDIA/nvidia-docker)
+- Any NVIDIA GPU supported by TensorFlow or PyTorch
+
+## Supported Models
+
+| model | framework | accuracy | dataset | model link | model source | precision | notes |
+| ----- | --------- | -------- | ------- | ---------- | ------------ | --------- | ----- |
+| BERT-Large | TensorFlow | f1_score=90.874% | SQuAD v1.1 validation set | [from zenodo](https://zenodo.org/record/3733868) [from zenodo](https://zenodo.org/record/3939747) | [BERT-Large](https://github.com/google-research/bert), trained with [NVIDIA DeepLearningExamples](https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow/LanguageModeling/BERT) | fp32 | |
+| BERT-Large | PyTorch | f1_score=90.874% | SQuAD v1.1 validation set | [from zenodo](https://zenodo.org/record/3733896) | [BERT-Large](https://github.com/google-research/bert), trained with [NVIDIA DeepLearningExamples](https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow/LanguageModeling/BERT), converted with [bert_tf_to_pytorch.py](bert_tf_to_pytorch.py) | fp32 | |
+| BERT-Large | ONNX | f1_score=90.874% | SQuAD v1.1 validation set | [from zenodo](https://zenodo.org/record/3733910) | [BERT-Large](https://github.com/google-research/bert), trained with [NVIDIA DeepLearningExamples](https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow/LanguageModeling/BERT), converted with [bert_tf_to_pytorch.py](bert_tf_to_pytorch.py) | fp32 | |
+| BERT-Large | ONNX | f1_score=90.067% | SQuAD v1.1 validation set | [from zenodo](https://zenodo.org/record/3750364) | Fine-tuned based on the PyTorch model and converted with [bert_tf_to_pytorch.py](bert_tf_to_pytorch.py) | int8, symetrically per-tensor quantized without bias | See [MLPerf INT8 BERT Finetuning.pdf](MLPerf INT8 BERT Finetuning.pdf) for details about the fine-tuning process |
+| BERT-Large | PyTorch | f1_score=90.633% | SQuAD v1.1 validation set | [from zenodo](https://zenodo.org/record/4792496) | Fine-tuned based on [Huggingface bert-large-uncased pretrained model](https://huggingface.co/bert-large-uncased) | int8, symetrically per-tensor quantized without bias | See README.md at Zenodo link for details about the fine-tuning process |
+
+## Disclaimer
+This benchmark app is a reference implementation that is not meant to be the fastest implementation possible.
+
+## Commands
+
+Please run the following commands:
+
+- `make setup`: initialize submodule, download datasets, and download models.
+- `make build_docker`: build docker image.
+- `make launch_docker`: launch docker container with an interaction session.
+- `python3 run.py --backend=[tf|pytorch|onnxruntime|tf_estimator] --scenario=[Offline|SingleStream|MultiStream|Server] [--accuracy] [--quantized]`: run the harness inside the docker container. Performance or Accuracy results will be printed in console.
+
+* ENV variable `CM_MAX_NUM_THREADS` can be used to control the number of parallel threads issuing queries.
+
+## Details
+
+- SUT implementations are in [tf_SUT.py](tf_SUT.py), [tf_estimator_SUT.py](tf_estimator_SUT.py) and [pytorch_SUT.py](pytorch_SUT.py). QSL implementation is in [squad_QSL.py](squad_QSL.py).
+- The script [accuracy-squad.py](accuracy-squad.py) parses LoadGen accuracy log, post-processes it, and computes the accuracy.
+- Tokenization and detokenization (post-processing) are not included in the timed path.
+- The inputs to the SUT are `input_ids`, `input_make`, and `segment_ids`. The output from SUT is `start_logits` and `end_logits` concatenated together.
+- `max_seq_length` is 384.
+- The script [tf_freeze_bert.py] freezes the TensorFlow model into pb file.
+- The script [bert_tf_to_pytorch.py] converts the TensorFlow model into the PyTorch `BertForQuestionAnswering` module in [HuggingFace Transformers](https://github.com/huggingface/transformers) and also exports the model to [ONNX](https://github.com/onnx/onnx) format.
+
+## Loadgen over the Network 
+
+```
+pip install cm4mlops
+```
+
+The below CM command will launch the SUT server
+
+```
+cm run script --tags=generate-run-cmds,inference --model=bert-99 --backend=pytorch  \
+--mode=performance --device=cuda --quiet --test_query_count=1000 --network=sut
+```
+
+Once the SUT server is launched, the below command can be run on the loadgen node to do issue queries to the SUT nodes. In this command `-sut_servers` has just the localhost address - it can be changed to a comma-separated list of any hostname/IP in the network. 
+
+
+```
+cm run script --tags=generate-run-cmds,inference --model=bert-99 --backend=pytorch  --rerun \
+--mode=performance --device=cuda --quiet --test_query_count=1000  \
+--sut_servers,=http://localhost:8000 --network=lon
+```
+
+If you are not using CM, just add `--network=lon` along with your normal run command on the SUT side.
+On the loadgen node, add `--network=lon` option and `--sut_server <IP1> <IP2>` to the normal command to connect to SUT nodes at IP addresses IP1, IP2 etc. 
+
+Loadgen over the network works for `onnxruntime` and `pytorch` backends.
+
+## License
+
+Apache License 2.0

language/bert/accuracy-squad.py

+# coding=utf-8
+# Copyright 2021 Arm Limited and affiliates.
+# Copyright (c) 2020 NVIDIA CORPORATION. All rights reserved.
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import collections
+import json
+import math
+import os
+import subprocess
+import sys
+
+import numpy as np
+import pkg_resources
+import six
+from transformers import BertTokenizer
+
+# To support feature cache.
+import pickle
+
+sys.path.insert(0, os.path.dirname(__file__))
+
+installed = {pkg.key for pkg in pkg_resources.working_set}
+if "tensorflow" in installed:
+    import tensorflow
+
+    sys.path.insert(
+        0,
+        os.path.join(
+            os.path.dirname(__file__),
+            "DeepLearningExamples",
+            "TensorFlow",
+            "LanguageModeling",
+            "BERT",
+        ),
+    )
+elif "torch" in installed:
+    import torch
+
+    sys.path.insert(
+        0,
+        os.path.join(
+            os.path.dirname(__file__),
+            "DeepLearningExamples",
+            "PyTorch",
+            "LanguageModeling",
+            "BERT",
+        ),
+    )
+try:
+    import tokenization
+    from create_squad_data import convert_examples_to_features, read_squad_examples
+except ImportError:
+    raise Exception("Error importing local modules")
+
+
+max_seq_length = 384
+max_query_length = 64
+doc_stride = 128
+
+RawResult = collections.namedtuple(
+    "RawResult", ["unique_id", "start_logits", "end_logits"]
+)
+
+dtype_map = {
+    "int8": np.int8,
+    "int16": np.int16,
+    "int32": np.int32,
+    "int64": np.int64,
+    "float16": np.float16,
+    "float32": np.float32,
+    "float64": np.float64}
+
+
+def get_final_text(pred_text, orig_text, do_lower_case):
+    """Project the tokenized prediction back to the original text."""
+
+    # When we created the data, we kept track of the alignment between original
+    # (whitespace tokenized) tokens and our WordPiece tokenized tokens. So
+    # now `orig_text` contains the span of our original text corresponding to the
+    # span that we predicted.
+    #
+    # However, `orig_text` may contain extra characters that we don't want in
+    # our prediction.
+    #
+    # For example, let's say:
+    #     pred_text = steve smith
+    #     orig_text = Steve Smith's
+    #
+    # We don't want to return `orig_text` because it contains the extra "'s".
+    #
+    # We don't want to return `pred_text` because it's already been normalized
+    # (the SQuAD eval script also does punctuation stripping/lower casing but
+    # our tokenizer does additional normalization like stripping accent
+    # characters).
+    #
+    # What we really want to return is "Steve Smith".
+    #
+    # Therefore, we have to apply a semi-complicated alignment heruistic between
+    # `pred_text` and `orig_text` to get a character-to-charcter alignment. This
+    # can fail in certain cases in which case we just return `orig_text`.
+
+    def _strip_spaces(text):
+        ns_chars = []
+        ns_to_s_map = collections.OrderedDict()
+        for i, c in enumerate(text):
+            if c == " ":
+                continue
+            ns_to_s_map[len(ns_chars)] = i
+            ns_chars.append(c)
+        ns_text = "".join(ns_chars)
+        return (ns_text, ns_to_s_map)
+
+    # We first tokenize `orig_text`, strip whitespace from the result
+    # and `pred_text`, and check if they are the same length. If they are
+    # NOT the same length, the heuristic has failed. If they are the same
+    # length, we assume the characters are one-to-one aligned.
+    tokenizer = tokenization.BasicTokenizer(do_lower_case=do_lower_case)
+
+    tok_text = " ".join(tokenizer.tokenize(orig_text))
+
+    start_position = tok_text.find(pred_text)
+    if start_position == -1:
+        return orig_text
+    end_position = start_position + len(pred_text) - 1
+
+    (orig_ns_text, orig_ns_to_s_map) = _strip_spaces(orig_text)
+    (tok_ns_text, tok_ns_to_s_map) = _strip_spaces(tok_text)
+
+    if len(orig_ns_text) != len(tok_ns_text):
+        return orig_text
+
+    # We then project the characters in `pred_text` back to `orig_text` using
+    # the character-to-character alignment.
+    tok_s_to_ns_map = {}
+    for i, tok_index in six.iteritems(tok_ns_to_s_map):
+        tok_s_to_ns_map[tok_index] = i
+
+    orig_start_position = None
+    if start_position in tok_s_to_ns_map:
+        ns_start_position = tok_s_to_ns_map[start_position]
+        if ns_start_position in orig_ns_to_s_map:
+            orig_start_position = orig_ns_to_s_map[ns_start_position]
+
+    if orig_start_position is None:
+        return orig_text
+
+    orig_end_position = None
+    if end_position in tok_s_to_ns_map:
+        ns_end_position = tok_s_to_ns_map[end_position]
+        if ns_end_position in orig_ns_to_s_map:
+            orig_end_position = orig_ns_to_s_map[ns_end_position]
+
+    if orig_end_position is None:
+        return orig_text
+
+    output_text = orig_text[orig_start_position: (orig_end_position + 1)]
+    return output_text
+
+
+def _get_best_indexes(logits, n_best_size):
+    """Get the n-best logits from a list."""
+    index_and_score = sorted(
+        enumerate(logits),
+        key=lambda x: x[1],
+        reverse=True)
+
+    best_indexes = []
+    for i in range(len(index_and_score)):
+        if i >= n_best_size:
+            break
+        best_indexes.append(index_and_score[i][0])
+    return best_indexes
+
+
+def _compute_softmax(scores):
+    """Compute softmax probability over raw logits."""
+    if not scores:
+        return []
+
+    max_score = None
+    for score in scores:
+        if max_score is None or score > max_score:
+            max_score = score
+
+    exp_scores = []
+    total_sum = 0.0
+    for score in scores:
+        x = math.exp(score - max_score)
+        exp_scores.append(x)
+        total_sum += x
+
+    probs = []
+    for score in exp_scores:
+        probs.append(score / total_sum)
+    return probs
+
+
+def write_predictions(
+    all_examples,
+    all_features,
+    all_results,
+    n_best_size,
+    max_answer_length,
+    do_lower_case,
+    output_prediction_file,
+    max_examples=None,
+):
+    """Write final predictions to the json file and log-odds of null if needed."""
+    print("Writing predictions to: %s" % (output_prediction_file))
+
+    example_index_to_features = collections.defaultdict(list)
+    for feature in all_features:
+        example_index_to_features[feature.example_index].append(feature)
+
+    unique_id_to_result = {}
+    for result in all_results:
+        unique_id_to_result[result.unique_id] = result
+
+    _PrelimPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+        "PrelimPrediction",
+        ["feature_index", "start_index", "end_index", "start_logit", "end_logit"],
+    )
+
+    all_predictions = collections.OrderedDict()
+    all_nbest_json = collections.OrderedDict()
+    scores_diff_json = collections.OrderedDict()
+
+    for example_index, example in enumerate(all_examples):
+        if max_examples and example_index == max_examples:
+            break
+
+        features = example_index_to_features[example_index]
+
+        prelim_predictions = []
+        # keep track of the minimum score of null start+end of position 0
+        score_null = 1000000  # large and positive
+        min_null_feature_index = 0  # the paragraph slice with min mull score
+        null_start_logit = 0  # the start logit at the slice with min null score
+        null_end_logit = 0  # the end logit at the slice with min null score
+        for feature_index, feature in enumerate(features):
+            # FIX: During compliance/audit runs, we only generate a small subset of
+            # all entries from the dataset. As a result, sometimes dict retrieval
+            # fails because a key is missing.
+            # result = unique_id_to_result[feature.unique_id]
+            result = unique_id_to_result.get(feature.unique_id, None)
+            if result is None:
+                continue
+            start_indexes = _get_best_indexes(result.start_logits, n_best_size)
+            end_indexes = _get_best_indexes(result.end_logits, n_best_size)
+            # if we could have irrelevant answers, get the min score of
+            # irrelevant
+            for start_index in start_indexes:
+                for end_index in end_indexes:
+                    # We could hypothetically create invalid predictions, e.g., predict
+                    # that the start of the span is in the question. We throw out all
+                    # invalid predictions.
+                    if start_index >= len(feature.tokens):
+                        continue
+                    if end_index >= len(feature.tokens):
+                        continue
+                    if start_index not in feature.token_to_orig_map:
+                        continue
+                    if end_index not in feature.token_to_orig_map:
+                        continue
+                    if not feature.token_is_max_context.get(
+                            start_index, False):
+                        continue
+                    if end_index < start_index:
+                        continue
+                    length = end_index - start_index + 1
+                    if length > max_answer_length:
+                        continue
+                    prelim_predictions.append(
+                        _PrelimPrediction(
+                            feature_index=feature_index,
+                            start_index=start_index,
+                            end_index=end_index,
+                            start_logit=result.start_logits[start_index],
+                            end_logit=result.end_logits[end_index],
+                        )
+                    )
+
+        prelim_predictions = sorted(
+            prelim_predictions,
+            key=lambda x: (x.start_logit + x.end_logit),
+            reverse=True,
+        )
+
+        _NbestPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+            "NbestPrediction", ["text", "start_logit", "end_logit"]
+        )
+
+        seen_predictions = {}
+        nbest = []
+        for pred in prelim_predictions:
+            if len(nbest) >= n_best_size:
+                break
+            feature = features[pred.feature_index]
+            tok_tokens = feature.tokens[pred.start_index: (pred.end_index + 1)]
+            orig_doc_start = feature.token_to_orig_map[pred.start_index]
+            orig_doc_end = feature.token_to_orig_map[pred.end_index]
+            orig_tokens = example.doc_tokens[orig_doc_start: (
+                orig_doc_end + 1)]
+            tok_text = " ".join(tok_tokens)
+
+            # De-tokenize WordPieces that have been split off.
+            tok_text = tok_text.replace(" ##", "")
+            tok_text = tok_text.replace("##", "")
+
+            # Clean whitespace
+            tok_text = tok_text.strip()
+            tok_text = " ".join(tok_text.split())
+            orig_text = " ".join(orig_tokens)
+
+            final_text = get_final_text(tok_text, orig_text, do_lower_case)
+            if final_text in seen_predictions:
+                continue
+
+            seen_predictions[final_text] = True
+
+            nbest.append(
+                _NbestPrediction(
+                    text=final_text,
+                    start_logit=pred.start_logit,
+                    end_logit=pred.end_logit,
+                )
+            )
+
+        # In very rare edge cases we could have no valid predictions. So we
+        # just create a nonce prediction in this case to avoid failure.
+        if not nbest:
+            nbest.append(
+                _NbestPrediction(
+                    text="empty",
+                    start_logit=0.0,
+                    end_logit=0.0))
+
+        assert len(nbest) >= 1
+
+        total_scores = []
+        best_non_null_entry = None
+        for entry in nbest:
+            total_scores.append(entry.start_logit + entry.end_logit)
+            if not best_non_null_entry:
+                if entry.text:
+                    best_non_null_entry = entry
+
+        probs = _compute_softmax(total_scores)
+
+        nbest_json = []
+        for i, entry in enumerate(nbest):
+            output = collections.OrderedDict()
+            output["text"] = entry.text
+            output["probability"] = probs[i]
+            output["start_logit"] = entry.start_logit
+            output["end_logit"] = entry.end_logit
+            nbest_json.append(output)
+
+        assert len(nbest_json) >= 1
+
+        all_predictions[example.qas_id] = nbest_json[0]["text"]
+
+    with open(output_prediction_file, "w") as writer:
+        writer.write(json.dumps(all_predictions, indent=4) + "\n")
+
+
+def load_loadgen_log(
+    log_path, eval_features, dtype=np.float32, output_transposed=False
+):
+    with open(log_path) as f:
+        predictions = json.load(f)
+
+    results = []
+    for prediction in predictions:
+        qsl_idx = prediction["qsl_idx"]
+        if output_transposed:
+            logits = np.frombuffer(bytes.fromhex(prediction["data"]), dtype).reshape(
+                2, -1
+            )
+            logits = np.transpose(logits)
+        else:
+            logits = np.frombuffer(bytes.fromhex(prediction["data"]), dtype).reshape(
+                -1, 2
+            )
+        # Pad logits to max_seq_length
+        seq_length = logits.shape[0]
+        start_logits = np.ones(max_seq_length) * -10000.0
+        end_logits = np.ones(max_seq_length) * -10000.0
+        start_logits[:seq_length] = logits[:, 0]
+        end_logits[:seq_length] = logits[:, 1]
+        results.append(
+            RawResult(
+                unique_id=eval_features[qsl_idx].unique_id,
+                start_logits=start_logits.tolist(),
+                end_logits=end_logits.tolist(),
+            )
+        )
+
+    return results
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--vocab_file",
+        default="build/data/bert_tf_v1_1_large_fp32_384_v2/vocab.txt",
+        help="Path to vocab.txt",
+    )
+    parser.add_argument(
+        "--val_data", default="build/data/dev-v1.1.json", help="Path to validation data"
+    )
+    parser.add_argument(
+        "--log_file",
+        default="build/logs/mlperf_log_accuracy.json",
+        help="Path to LoadGen accuracy log",
+    )
+    parser.add_argument(
+        "--out_file",
+        default="build/result/predictions.json",
+        help="Path to output predictions file",
+    )
+    parser.add_argument(
+        "--features_cache_file",
+        default="eval_features.pickle",
+        help="Path to features' cache file",
+    )
+    parser.add_argument(
+        "--output_transposed", action="store_true", help="Transpose the output"
+    )
+    parser.add_argument(
+        "--output_dtype",
+        default="float32",
+        choices=dtype_map.keys(),
+        help="Output data type",
+    )
+    parser.add_argument(
+        "--max_examples",
+        type=int,
+        help="Maximum number of examples to consider (not limited by default)",
+    )
+    args = parser.parse_args()
+
+    output_dtype = dtype_map[args.output_dtype]
+
+    print("Reading examples...")
+    eval_examples = read_squad_examples(
+        input_file=args.val_data, is_training=False, version_2_with_negative=False
+    )
+
+    eval_features = []
+    # Load features if cached, convert from examples otherwise.
+    cache_path = args.features_cache_file
+    if os.path.exists(cache_path):
+        print("Loading cached features from '%s'..." % cache_path)
+        with open(cache_path, "rb") as cache_file:
+            eval_features = pickle.load(cache_file)
+    else:
+        print(
+            "No cached features at '%s'... converting from examples..." %
+            cache_path)
+
+        print("Creating tokenizer...")
+        tokenizer = BertTokenizer(args.vocab_file)
+
+        print("Converting examples to features...")
+
+        def append_feature(feature):
+            eval_features.append(feature)
+
+        convert_examples_to_features(
+            examples=eval_examples,
+            tokenizer=tokenizer,
+            max_seq_length=max_seq_length,
+            doc_stride=doc_stride,
+            max_query_length=max_query_length,
+            is_training=False,
+            output_fn=append_feature,
+            verbose_logging=False,
+        )
+
+        print("Caching features at '%s'..." % cache_path)
+        with open(cache_path, "wb") as cache_file:
+            pickle.dump(eval_features, cache_file)
+
+    print("Loading LoadGen logs...")
+    results = load_loadgen_log(
+        args.log_file, eval_features, output_dtype, args.output_transposed
+    )
+
+    print("Post-processing predictions...")
+    write_predictions(
+        eval_examples,
+        eval_features,
+        results,
+        20,
+        30,
+        True,
+        args.out_file,
+        args.max_examples,
+    )
+
+    print("Evaluating predictions...")
+    cmd = "python3 {:}/evaluate_v1.1.py {:} {:} {}".format(
+        os.path.dirname(os.path.abspath(__file__)),
+        args.val_data,
+        args.out_file,
+        "--max_examples {}".format(
+            args.max_examples) if args.max_examples else "",
+    )
+    subprocess.check_call(cmd, shell=True)
+
+
+if __name__ == "__main__":
+    main()

language/bert/bert_QDL.py

+# Copyright 2023 MLCommons. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# =============================================================================
+
+
+from time import sleep
+import squad_QSL
+import mlperf_loadgen as lg
+import numpy as np
+import threading
+import requests
+import array
+import time
+import os
+import sys
+
+sys.path.insert(0, os.getcwd())
+
+
+class bert_QDL:
+    """QDL acting as a proxy to the SUT.
+    This QDL communicates with the SUT via HTTP.
+    It uses two endpoints to communicate with the SUT:
+    - /predict/ : Send a query to the SUT and get a response.
+    - /getname/ : Get the name of the SUT. Send a getname to the SUT and get a response.
+    """
+
+    def __init__(self, qsl: squad_QSL.SQuAD_v1_QSL, sut_server_addr: list):
+        """
+        Constructor for the QDL.
+        Args:
+            qsl: The QSL to use.
+            sut_server_addr: A list of addresses of the SUT.
+        """
+        self.qsl = qsl
+        self.quantized = False
+
+        # Construct QDL from the python binding
+        self.qdl = lg.ConstructQDL(
+            self.issue_query, self.flush_queries, self.client_get_name
+        )
+        self.sut_server_addr = sut_server_addr
+        self.num_nodes = len(sut_server_addr)
+
+        # For round robin between the SUTs:
+        self.next_sut_id = 0
+        self.lock = threading.Lock()
+
+    def issue_query(self, query_samples):
+        """Process the query to send to the SUT"""
+        threading.Thread(
+            target=self.process_query_async,
+            args=[query_samples]).start()
+
+    def flush_queries(self):
+        """Flush the queries. Dummy implementation."""
+        pass
+
+    def process_query_async(self, query_samples):
+        """
+        This function is called by the Loadgen in a separate thread.
+        It is responsible for
+            1. Creating a query for the SUT, by reading the features from the QSL.
+            2. Sending the query to the SUT.
+            3. Waiting for the response from the SUT.
+            4. Deserializing the response.
+            5. Calling mlperf_loadgen.QuerySamplesComplete(query_samples, response)
+        Args:
+            query_samples: A list of QuerySample objects.
+        """
+
+        max_num_threads = int(
+            os.environ.get(
+                "CM_MAX_NUM_THREADS",
+                os.cpu_count()))
+
+        for i in range(len(query_samples)):
+            eval_features = self.qsl.get_features(query_samples[i].index)
+            encoded_eval_features = {
+                "input_ids": eval_features.input_ids,
+                "input_mask": eval_features.input_mask,
+                "segment_ids": eval_features.segment_ids,
+            }
+            n = threading.active_count()
+            while n >= max_num_threads:
+                sleep(0.0001)
+                n = threading.active_count()
+            threading.Thread(
+                target=self.client_predict_worker,
+                args=[encoded_eval_features, query_samples[i].id],
+            ).start()
+
+    def get_sut_id_round_robin(self):
+        """Get the SUT id in round robin."""
+        with self.lock:
+            res = self.next_sut_id
+            self.next_sut_id = (self.next_sut_id + 1) % self.num_nodes
+        return res
+
+    def client_predict_worker(self, query, query_id):
+        """Serialize the query, send it to the SUT in round robin, and return the deserialized response."""
+        url = "{}/predict/".format(
+            self.sut_server_addr[self.get_sut_id_round_robin()])
+        responses = []
+        response = requests.post(url, json={"query": query})
+        output = response.json()["result"]
+        output = np.array(output).astype(np.float32)
+        response_array = array.array("B", output.tobytes())
+        bi = response_array.buffer_info()
+
+        responses.append(lg.QuerySampleResponse(query_id, bi[0], bi[1]))
+        lg.QuerySamplesComplete(responses)
+
+    def client_get_name(self):
+        """Get the name of the SUT from ALL the SUTS."""
+        if len(self.sut_server_addr) == 1:
+            return requests.post(
+                f"{self.sut_server_addr[0]}/getname/").json()["name"]
+
+        sut_names = [
+            requests.post(f"{addr}/getname/").json()["name"]
+            for addr in self.sut_server_addr
+        ]
+        return "Multi-node SUT: " + ", ".join(sut_names)
+
+    def __del__(self):
+        lg.DestroyQDL(self.qdl)

language/bert/bert_config.json

+{
+  "attention_probs_dropout_prob": 0.1,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "max_position_embeddings": 512,
+  "num_attention_heads": 16,
+  "num_hidden_layers": 24,
+  "type_vocab_size": 2,
+  "vocab_size": 30522
+}

language/bert/bert_tf_to_pytorch.py

+# coding=utf-8
+# Copyright (c) 2020 NVIDIA CORPORATION. All rights reserved.
+# Copyright 2018 The HuggingFace Inc. team.
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import collections
+import json
+import math
+import os
+import random
+import re
+import shutil
+import sys
+import time
+
+sys.path.insert(
+    0,
+    os.path.join(
+        os.getcwd(), "DeepLearningExamples", "TensorFlow", "LanguageModeling", "BERT"
+    ),
+)
+sys.path.insert(0, os.getcwd())
+
+try:
+    import tensorflow as tf
+    from transformers import BertConfig, BertTokenizer, BertForQuestionAnswering
+    import torch
+    import numpy as np
+except ImportError:
+    raise Exception("Error importing local modules")
+
+
+def load_from_tf(config, tf_path):
+    model = BertForQuestionAnswering(config)
+    model.classifier = model.qa_outputs
+
+    # This part is copied from HuggingFace Transformers with a fix to bypass
+    # an error
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    arrays = []
+    for name, shape in init_vars:
+        # print("Loading TF weight {} with shape {}".format(name, shape))
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        arrays.append(array)
+
+    for name, array in zip(names, arrays):
+        name = name.split("/")
+        # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
+        # which are not required for using pretrained model
+        if any(n in ["adam_v", "adam_m", "global_step"] for n in name):
+            print("Skipping {}".format("/".join(name)))
+            continue
+        pointer = model
+        for m_name in name:
+            if re.fullmatch(r"[A-Za-z]+_\d+", m_name):
+                scope_names = re.split(r"_(\d+)", m_name)
+            else:
+                scope_names = [m_name]
+            if scope_names[0] == "kernel" or scope_names[0] == "gamma":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "output_bias" or scope_names[0] == "beta":
+                pointer = getattr(pointer, "bias")
+            elif scope_names[0] == "output_weights":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "squad":
+                pointer = getattr(
+                    pointer, "classifier"
+                )  # This line is causing the issue
+            else:
+                try:
+                    pointer = getattr(pointer, scope_names[0])
+                except AttributeError:
+                    print("Skipping {}".format("/".join(name)))
+                    continue
+            if len(scope_names) >= 2:
+                num = int(scope_names[1])
+                pointer = pointer[num]
+        if m_name[-11:] == "_embeddings":
+            pointer = getattr(pointer, "weight")
+        elif m_name == "kernel":
+            array = np.transpose(array)
+        try:
+            assert pointer.shape == array.shape
+        except AssertionError as e:
+            e.args += (pointer.shape, array.shape)
+            raise
+        print("Initialize PyTorch weight {}".format(name))
+        pointer.data = torch.from_numpy(array)
+
+    model.qa_outputs = model.classifier
+    del model.classifier
+    return model
+
+
+def save_to_onnx(model):
+    tokenizer = BertTokenizer.from_pretrained(
+        "bert-large-uncased-whole-word-masking-finetuned-squad"
+    )
+    model.eval()
+
+    dummy_input = torch.ones((1, 384), dtype=torch.int64)
+    torch.onnx.export(
+        model,
+        (dummy_input, dummy_input, dummy_input),
+        "build/data/bert_tf_v1_1_large_fp32_384_v2/model.onnx",
+        verbose=True,
+        input_names=["input_ids", "input_mask", "segment_ids"],
+        output_names=["output_start_logits", "output_end_logits"],
+        opset_version=11,
+        dynamic_axes=(
+            {
+                "input_ids": {0: "batch_size"},
+                "input_mask": {0: "batch_size"},
+                "segment_ids": {0: "batch_size"},
+                "output_start_logits": {0: "batch_size"},
+                "output_end_logits": {0: "batch_size"},
+            }
+        ),
+    )
+
+
+def main():
+    with open("build/data/bert_tf_v1_1_large_fp32_384_v2/bert_config.json") as f:
+        config_json = json.load(f)
+
+    config = BertConfig(
+        attention_probs_dropout_prob=config_json["attention_probs_dropout_prob"],
+        hidden_act=config_json["hidden_act"],
+        hidden_dropout_prob=config_json["hidden_dropout_prob"],
+        hidden_size=config_json["hidden_size"],
+        initializer_range=config_json["initializer_range"],
+        intermediate_size=config_json["intermediate_size"],
+        max_position_embeddings=config_json["max_position_embeddings"],
+        num_attention_heads=config_json["num_attention_heads"],
+        num_hidden_layers=config_json["num_hidden_layers"],
+        type_vocab_size=config_json["type_vocab_size"],
+        vocab_size=config_json["vocab_size"],
+    )
+
+    model = load_from_tf(
+        config, "build/data/bert_tf_v1_1_large_fp32_384_v2/model.ckpt-5474"
+    )
+    torch.save(
+        model.state_dict(), "build/data/bert_tf_v1_1_large_fp32_384_v2/model.pytorch"
+    )
+    save_to_onnx(model)
+
+
+if __name__ == "__main__":
+    main()

language/bert/create_squad_data.py

+# Copyright 2021 Arm Limited and affiliates.
+# Copyright (c) 2019 NVIDIA CORPORATION. All rights reserved.
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This file is identical to DeepLearningExamples/TensorFlow/LanguageModeling/BERT/utils/create_squad_data.py
+# except that the dependency on horovod is removed.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import collections
+import json
+import tokenization
+import six
+
+
+class SquadExample(object):
+    """A single training/test example for simple sequence classification.
+
+    For examples without an answer, the start and end position are -1.
+    """
+
+    def __init__(
+        self,
+        qas_id,
+        question_text,
+        doc_tokens,
+        orig_answer_text=None,
+        start_position=None,
+        end_position=None,
+        is_impossible=False,
+    ):
+        self.qas_id = qas_id
+        self.question_text = question_text
+        self.doc_tokens = doc_tokens
+        self.orig_answer_text = orig_answer_text
+        self.start_position = start_position
+        self.end_position = end_position
+        self.is_impossible = is_impossible
+
+    def __str__(self):
+        return self.__repr__()
+
+    def __repr__(self):
+        s = ""
+        s += "qas_id: %s" % (tokenization.printable_text(self.qas_id))
+        s += ", question_text: %s" % (
+            tokenization.printable_text(self.question_text))
+        s += ", doc_tokens: [%s]" % (" ".join(self.doc_tokens))
+        if self.start_position:
+            s += ", start_position: %d" % (self.start_position)
+        if self.start_position:
+            s += ", end_position: %d" % (self.end_position)
+        if self.start_position:
+            s += ", is_impossible: %r" % (self.is_impossible)
+        return s
+
+
+class InputFeatures(object):
+    """A single set of features of data."""
+
+    def __init__(
+        self,
+        unique_id,
+        example_index,
+        doc_span_index,
+        tokens,
+        token_to_orig_map,
+        token_is_max_context,
+        input_ids,
+        input_mask,
+        segment_ids,
+        start_position=None,
+        end_position=None,
+        is_impossible=None,
+    ):
+        self.unique_id = unique_id
+        self.example_index = example_index
+        self.doc_span_index = doc_span_index
+        self.tokens = tokens
+        self.token_to_orig_map = token_to_orig_map
+        self.token_is_max_context = token_is_max_context
+        self.input_ids = input_ids
+        self.input_mask = input_mask
+        self.segment_ids = segment_ids
+        self.start_position = start_position
+        self.end_position = end_position
+        self.is_impossible = is_impossible
+
+
+def read_squad_examples(input_file, is_training,
+                        version_2_with_negative=False):
+    """Read a SQuAD json file into a list of SquadExample."""
+    with open(input_file) as reader:
+        input_data = json.load(reader)["data"]
+
+    def is_whitespace(c):
+        if c == " " or c == "\t" or c == "\r" or c == "\n" or ord(c) == 0x202F:
+            return True
+        return False
+
+    examples = []
+    for entry in input_data:
+        for paragraph in entry["paragraphs"]:
+            paragraph_text = paragraph["context"]
+            doc_tokens = []
+            char_to_word_offset = []
+            prev_is_whitespace = True
+            for c in paragraph_text:
+                if is_whitespace(c):
+                    prev_is_whitespace = True
+                else:
+                    if prev_is_whitespace:
+                        doc_tokens.append(c)
+                    else:
+                        doc_tokens[-1] += c
+                    prev_is_whitespace = False
+                char_to_word_offset.append(len(doc_tokens) - 1)
+
+            for qa in paragraph["qas"]:
+                qas_id = qa["id"]
+                question_text = qa["question"]
+                start_position = None
+                end_position = None
+                orig_answer_text = None
+                is_impossible = False
+                if is_training:
+
+                    if version_2_with_negative:
+                        is_impossible = qa["is_impossible"]
+                    if (len(qa["answers"]) != 1) and (not is_impossible):
+                        raise ValueError(
+                            "For training, each question should have exactly 1 answer."
+                        )
+                    if not is_impossible:
+                        answer = qa["answers"][0]
+                        orig_answer_text = answer["text"]
+                        answer_offset = answer["answer_start"]
+                        answer_length = len(orig_answer_text)
+                        start_position = char_to_word_offset[answer_offset]
+                        end_position = char_to_word_offset[
+                            answer_offset + answer_length - 1
+                        ]
+                        # Only add answers where the text can be exactly recovered from the
+                        # document. If this CAN'T happen it's likely due to weird Unicode
+                        # stuff so we will just skip the example.
+                        #
+                        # Note that this means for training mode, every example is NOT
+                        # guaranteed to be preserved.
+                        actual_text = " ".join(
+                            doc_tokens[start_position: (end_position + 1)]
+                        )
+                        cleaned_answer_text = " ".join(
+                            tokenization.whitespace_tokenize(orig_answer_text)
+                        )
+                        if actual_text.find(cleaned_answer_text) == -1:
+                            print(
+                                "Could not find answer: '%s' vs. '%s'",
+                                actual_text,
+                                cleaned_answer_text,
+                            )
+                            continue
+                    else:
+                        start_position = -1
+                        end_position = -1
+                        orig_answer_text = ""
+
+                example = SquadExample(
+                    qas_id=qas_id,
+                    question_text=question_text,
+                    doc_tokens=doc_tokens,
+                    orig_answer_text=orig_answer_text,
+                    start_position=start_position,
+                    end_position=end_position,
+                    is_impossible=is_impossible,
+                )
+                examples.append(example)
+
+    return examples
+
+
+def _check_is_max_context(doc_spans, cur_span_index, position):
+    """Check if this is the 'max context' doc span for the token."""
+
+    # Because of the sliding window approach taken to scoring documents, a single
+    # token can appear in multiple documents. E.g.
+    #  Doc: the man went to the store and bought a gallon of milk
+    #  Span A: the man went to the
+    #  Span B: to the store and bought
+    #  Span C: and bought a gallon of
+    #  ...
+    #
+    # Now the word 'bought' will have two scores from spans B and C. We only
+    # want to consider the score with "maximum context", which we define as
+    # the *minimum* of its left and right context (the *sum* of left and
+    # right context will always be the same, of course).
+    #
+    # In the example the maximum context for 'bought' would be span C since
+    # it has 1 left context and 3 right context, while span B has 4 left context
+    # and 0 right context.
+    best_score = None
+    best_span_index = None
+    for span_index, doc_span in enumerate(doc_spans):
+        end = doc_span.start + doc_span.length - 1
+        if position < doc_span.start:
+            continue
+        if position > end:
+            continue
+        num_left_context = position - doc_span.start
+        num_right_context = end - position
+        score = min(num_left_context, num_right_context) + \
+            0.01 * doc_span.length
+        if best_score is None or score > best_score:
+            best_score = score
+            best_span_index = span_index
+
+    return cur_span_index == best_span_index
+
+
+def _improve_answer_span(
+    doc_tokens, input_start, input_end, tokenizer, orig_answer_text
+):
+    """Returns tokenized answer spans that better match the annotated answer."""
+
+    # The SQuAD annotations are character based. We first project them to
+    # whitespace-tokenized words. But then after WordPiece tokenization, we can
+    # often find a "better match". For example:
+    #
+    #   Question: What year was John Smith born?
+    #   Context: The leader was John Smith (1895-1943).
+    #   Answer: 1895
+    #
+    # The original whitespace-tokenized answer will be "(1895-1943).". However
+    # after tokenization, our tokens will be "( 1895 - 1943 ) .". So we can match
+    # the exact answer, 1895.
+    #
+    # However, this is not always possible. Consider the following:
+    #
+    #   Question: What country is the top exporter of electornics?
+    #   Context: The Japanese electronics industry is the lagest in the world.
+    #   Answer: Japan
+    #
+    # In this case, the annotator chose "Japan" as a character sub-span of
+    # the word "Japanese". Since our WordPiece tokenizer does not split
+    # "Japanese", we just use "Japanese" as the annotation. This is fairly rare
+    # in SQuAD, but does happen.
+    tok_answer_text = " ".join(tokenizer.tokenize(orig_answer_text))
+
+    for new_start in range(input_start, input_end + 1):
+        for new_end in range(input_end, new_start - 1, -1):
+            text_span = " ".join(doc_tokens[new_start: (new_end + 1)])
+            if text_span == tok_answer_text:
+                return (new_start, new_end)
+
+    return (input_start, input_end)
+
+
+def convert_examples_to_features(
+    examples,
+    tokenizer,
+    max_seq_length,
+    doc_stride,
+    max_query_length,
+    is_training,
+    output_fn,
+    verbose_logging=False,
+):
+    """Loads a data file into a list of `InputBatch`s."""
+
+    unique_id = 1000000000
+
+    for example_index, example in enumerate(examples):
+        query_tokens = tokenizer.tokenize(example.question_text)
+
+        if len(query_tokens) > max_query_length:
+            query_tokens = query_tokens[0:max_query_length]
+
+        tok_to_orig_index = []
+        orig_to_tok_index = []
+        all_doc_tokens = []
+        for i, token in enumerate(example.doc_tokens):
+            orig_to_tok_index.append(len(all_doc_tokens))
+            sub_tokens = tokenizer.tokenize(token)
+            for sub_token in sub_tokens:
+                tok_to_orig_index.append(i)
+                all_doc_tokens.append(sub_token)
+
+        tok_start_position = None
+        tok_end_position = None
+        if is_training and example.is_impossible:
+            tok_start_position = -1
+            tok_end_position = -1
+        if is_training and not example.is_impossible:
+            tok_start_position = orig_to_tok_index[example.start_position]
+            if example.end_position < len(example.doc_tokens) - 1:
+                tok_end_position = orig_to_tok_index[example.end_position + 1] - 1
+            else:
+                tok_end_position = len(all_doc_tokens) - 1
+            (tok_start_position, tok_end_position) = _improve_answer_span(
+                all_doc_tokens,
+                tok_start_position,
+                tok_end_position,
+                tokenizer,
+                example.orig_answer_text,
+            )
+
+        # The -3 accounts for [CLS], [SEP] and [SEP]
+        max_tokens_for_doc = max_seq_length - len(query_tokens) - 3
+
+        # We can have documents that are longer than the maximum sequence length.
+        # To deal with this we do a sliding window approach, where we take chunks
+        # of the up to our max length with a stride of `doc_stride`.
+        _DocSpan = collections.namedtuple(  # pylint: disable=invalid-name
+            "DocSpan", ["start", "length"]
+        )
+        doc_spans = []
+        start_offset = 0
+        while start_offset < len(all_doc_tokens):
+            length = len(all_doc_tokens) - start_offset
+            if length > max_tokens_for_doc:
+                length = max_tokens_for_doc
+            doc_spans.append(_DocSpan(start=start_offset, length=length))
+            if start_offset + length == len(all_doc_tokens):
+                break
+            start_offset += min(length, doc_stride)
+
+        for doc_span_index, doc_span in enumerate(doc_spans):
+            tokens = []
+            token_to_orig_map = {}
+            token_is_max_context = {}
+            segment_ids = []
+            tokens.append("[CLS]")
+            segment_ids.append(0)
+            for token in query_tokens:
+                tokens.append(token)
+                segment_ids.append(0)
+            tokens.append("[SEP]")
+            segment_ids.append(0)
+
+            for i in range(doc_span.length):
+                split_token_index = doc_span.start + i
+                token_to_orig_map[len(
+                    tokens)] = tok_to_orig_index[split_token_index]
+
+                is_max_context = _check_is_max_context(
+                    doc_spans, doc_span_index, split_token_index
+                )
+                token_is_max_context[len(tokens)] = is_max_context
+                tokens.append(all_doc_tokens[split_token_index])
+                segment_ids.append(1)
+            tokens.append("[SEP]")
+            segment_ids.append(1)
+
+            input_ids = tokenizer.convert_tokens_to_ids(tokens)
+
+            # The mask has 1 for real tokens and 0 for padding tokens. Only real
+            # tokens are attended to.
+            input_mask = [1] * len(input_ids)
+
+            # Zero-pad up to the sequence length.
+            while len(input_ids) < max_seq_length:
+                input_ids.append(0)
+                input_mask.append(0)
+                segment_ids.append(0)
+
+            assert len(input_ids) == max_seq_length
+            assert len(input_mask) == max_seq_length
+            assert len(segment_ids) == max_seq_length
+
+            start_position = None
+            end_position = None
+            if is_training and not example.is_impossible:
+                # For training, if our document chunk does not contain an annotation
+                # we throw it out, since there is nothing to predict.
+                doc_start = doc_span.start
+                doc_end = doc_span.start + doc_span.length - 1
+                out_of_span = False
+                if not (
+                    tok_start_position >= doc_start and tok_end_position <= doc_end
+                ):
+                    out_of_span = True
+                if out_of_span:
+                    start_position = 0
+                    end_position = 0
+                else:
+                    doc_offset = len(query_tokens) + 2
+                    start_position = tok_start_position - doc_start + doc_offset
+                    end_position = tok_end_position - doc_start + doc_offset
+
+            if is_training and example.is_impossible:
+                start_position = 0
+                end_position = 0
+
+            if verbose_logging and example_index < 20:
+                print("*** Example ***")
+                print("unique_id: %s" % (unique_id))
+                print("example_index: %s" % (example_index))
+                print("doc_span_index: %s" % (doc_span_index))
+                print(
+                    "tokens: %s"
+                    % " ".join([tokenization.printable_text(x) for x in tokens])
+                )
+                print(
+                    "token_to_orig_map: %s"
+                    % " ".join(
+                        [
+                            "%d:%d" % (x, y)
+                            for (x, y) in six.iteritems(token_to_orig_map)
+                        ]
+                    )
+                )
+                print(
+                    "token_is_max_context: %s"
+                    % " ".join(
+                        [
+                            "%d:%s" % (x, y)
+                            for (x, y) in six.iteritems(token_is_max_context)
+                        ]
+                    )
+                )
+                print("input_ids: %s" % " ".join([str(x) for x in input_ids]))
+                print("input_mask: %s" %
+                      " ".join([str(x) for x in input_mask]))
+                print("segment_ids: %s" %
+                      " ".join([str(x) for x in segment_ids]))
+                if is_training and example.is_impossible:
+                    print("impossible example")
+                if is_training and not example.is_impossible:
+                    answer_text = " ".join(
+                        tokens[start_position: (end_position + 1)])
+                    print("start_position: %d" % (start_position))
+                    print("end_position: %d" % (end_position))
+                    print(
+                        "answer: %s" %
+                        (tokenization.printable_text(answer_text)))
+
+            feature = InputFeatures(
+                unique_id=unique_id,
+                example_index=example_index,
+                doc_span_index=doc_span_index,
+                tokens=tokens,
+                token_to_orig_map=token_to_orig_map,
+                token_is_max_context=token_is_max_context,
+                input_ids=input_ids,
+                input_mask=input_mask,
+                segment_ids=segment_ids,
+                start_position=start_position,
+                end_position=end_position,
+                is_impossible=example.is_impossible,
+            )
+
+            # Run callback
+            output_fn(feature)
+
+            unique_id += 1

language/bert/evaluate_v1.1.py

+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# Source:
+# https://github.com/allenai/bi-att-flow/blob/master/squad/evaluate-v1.1.py
+
+""" Official evaluation script for v1.1 of the SQuAD dataset. """
+from __future__ import print_function
+from collections import Counter
+import string
+import re
+import argparse
+import json
+import sys
+
+
+def normalize_answer(s):
+    """Lower text and remove punctuation, articles and extra whitespace."""
+
+    def remove_articles(text):
+        return re.sub(r"\b(a|an|the)\b", " ", text)
+
+    def white_space_fix(text):
+        return " ".join(text.split())
+
+    def remove_punc(text):
+        exclude = set(string.punctuation)
+        return "".join(ch for ch in text if ch not in exclude)
+
+    def lower(text):
+        return text.lower()
+
+    return white_space_fix(remove_articles(remove_punc(lower(s))))
+
+
+def f1_score(prediction, ground_truth):
+    prediction_tokens = normalize_answer(prediction).split()
+    ground_truth_tokens = normalize_answer(ground_truth).split()
+    common = Counter(prediction_tokens) & Counter(ground_truth_tokens)
+    num_same = sum(common.values())
+    if num_same == 0:
+        return 0
+    precision = 1.0 * num_same / len(prediction_tokens)
+    recall = 1.0 * num_same / len(ground_truth_tokens)
+    f1 = (2 * precision * recall) / (precision + recall)
+    return f1
+
+
+def exact_match_score(prediction, ground_truth):
+    return normalize_answer(prediction) == normalize_answer(ground_truth)
+
+
+def metric_max_over_ground_truths(metric_fn, prediction, ground_truths):
+    scores_for_ground_truths = []
+    for ground_truth in ground_truths:
+        score = metric_fn(prediction, ground_truth)
+        scores_for_ground_truths.append(score)
+    return max(scores_for_ground_truths)
+
+
+def evaluate(dataset, predictions, max_examples=None):
+    f1 = exact_match = total = 0
+    for article in dataset:
+        if max_examples and max_examples == total:
+            break
+        for paragraph in article["paragraphs"]:
+            if max_examples and max_examples == total:
+                break
+            for qa in paragraph["qas"]:
+                total += 1
+                if max_examples and max_examples == total:
+                    break
+
+                if qa["id"] not in predictions:
+                    message = (
+                        "Unanswered question " +
+                        qa["id"] + " will receive score 0."
+                    )
+                    print(message, file=sys.stderr)
+                    continue
+                ground_truths = list(map(lambda x: x["text"], qa["answers"]))
+                prediction = predictions[qa["id"]]
+                exact_match += metric_max_over_ground_truths(
+                    exact_match_score, prediction, ground_truths
+                )
+                f1 += metric_max_over_ground_truths(
+                    f1_score, prediction, ground_truths)
+
+    exact_match = 100.0 * exact_match / total
+    f1 = 100.0 * f1 / total
+
+    return {"exact_match": exact_match, "f1": f1}
+
+
+if __name__ == "__main__":
+    expected_version = "1.1"
+    parser = argparse.ArgumentParser(
+        description="Evaluation for SQuAD " + expected_version
+    )
+    parser.add_argument("dataset_file", help="Dataset file")
+    parser.add_argument("prediction_file", help="Prediction File")
+    parser.add_argument(
+        "--max_examples",
+        type=int,
+        help="Maximum number of examples to consider (not limited by default)",
+    )
+    args = parser.parse_args()
+    with open(args.dataset_file) as dataset_file:
+        dataset_json = json.load(dataset_file)
+        if dataset_json["version"] != expected_version:
+            print(
+                "Evaluation expects v-"
+                + expected_version
+                + ", but got dataset with v-"
+                + dataset_json["version"],
+                file=sys.stderr,
+            )
+        dataset = dataset_json["data"]
+    with open(args.prediction_file) as prediction_file:
+        predictions = json.load(prediction_file)
+    print(json.dumps(evaluate(dataset, predictions, args.max_examples)))

language/bert/network_LON.py

+# Copyright 2023 MLCommons. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# =============================================================================
+
+
+import bert_QDL
+import mlperf_loadgen as lg
+import squad_QSL
+from absl import app
+import sys
+import os
+
+sys.path.insert(0, os.getcwd())
+
+
+def set_args(
+    argv,
+    g_settings,
+    g_log_settings,
+    g_audit_conf,
+    g_sut_server,
+    g_backend,
+    g_total_count_override=None,
+    g_perf_count_override=None,
+):
+
+    global settings, log_settings, audit_conf, sut_server, total_count_override, perf_count_override, backend
+    sys.argv = sys.argv[0:1]
+    settings = g_settings
+    log_settings = g_log_settings
+    audit_conf = g_audit_conf
+    sut_server = g_sut_server
+    total_count_override = g_total_count_override
+    perf_count_override = g_perf_count_override
+    backend = g_backend
+
+
+def main(argv):
+    qsl = squad_QSL.get_squad_QSL(total_count_override, perf_count_override)
+    qdl = bert_QDL.bert_QDL(qsl, sut_server_addr=sut_server)
+
+    lg.StartTestWithLogSettings(
+        qdl.qdl,
+        qsl.qsl,
+        settings,
+        log_settings,
+        audit_conf)
+
+
+if __name__ == "__main__":
+    app.run(main)

language/bert/onnxruntime_SUT.py

+# coding=utf-8
+# Copyright (c) 2020 NVIDIA CORPORATION. All rights reserved.
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from time import sleep
+from squad_QSL import get_squad_QSL
+from transformers import BertConfig, BertForQuestionAnswering
+import onnxruntime
+import numpy as np
+import mlperf_loadgen as lg
+import threading
+import array
+import json
+import os
+import sys
+
+sys.path.insert(0, os.getcwd())
+
+
+class BERT_ONNXRuntime_SUT:
+    def __init__(self, args):
+        self.profile = args.profile
+        self.network = args.network
+        self.options = onnxruntime.SessionOptions()
+        self.options.enable_profiling = args.profile
+
+        print("Loading ONNX model...")
+        self.quantized = args.quantized
+
+        model_path = os.environ.get("ML_MODEL_FILE_WITH_PATH")
+        if not model_path:
+            if self.quantized:
+                model_path = "build/data/bert_tf_v1_1_large_fp32_384_v2/bert_large_v1_1_fake_quant.onnx"
+            else:
+                model_path = "build/data/bert_tf_v1_1_large_fp32_384_v2/model.onnx"
+        if len(onnxruntime.get_all_providers()) > 1 and os.environ.get(
+            "USE_GPU", "yes"
+        ).lower() not in ["0", "false", "off", "no"]:
+            preferred_execution_provider = os.environ.get(
+                "ONNXRUNTIME_PREFERRED_EXECUTION_PROVIDER", "CUDAExecutionProvider"
+            )
+            self.sess = onnxruntime.InferenceSession(
+                model_path, self.options, providers=[
+                    preferred_execution_provider]
+            )
+        else:
+            self.sess = onnxruntime.InferenceSession(
+                model_path, self.options, providers=["CPUExecutionProvider"]
+            )
+
+        print("Constructing SUT...")
+        self.sut = lg.ConstructSUT(self.issue_queries, self.flush_queries)
+        print("Finished constructing SUT.")
+
+        self.qsl = get_squad_QSL(args.max_examples)
+
+    def issue_queries(self, query_samples):
+        max_num_threads = int(
+            os.environ.get(
+                "CM_MAX_NUM_THREADS",
+                os.cpu_count()))
+
+        for i in range(len(query_samples)):
+            eval_features = self.qsl.get_features(query_samples[i].index)
+            n = threading.active_count()
+            while n >= max_num_threads:
+                # sleep(0.01)
+                n = threading.active_count()
+            threading.Thread(
+                target=self.process_sample, args=[
+                    eval_features, query_samples[i].id]
+            ).start()
+
+    def process_sample(self, eval_features, query_id=None):
+        """For Loadgen over the network"""
+        if self.network == "sut":
+            input_ids = eval_features["input_ids"]
+            input_mask = eval_features["input_mask"]
+            segment_ids = eval_features["segment_ids"]
+        else:
+            input_ids = eval_features.input_ids
+            input_mask = eval_features.input_mask
+            segment_ids = eval_features.segment_ids
+
+        if self.quantized:
+            fd = {
+                "input_ids": np.array(input_ids).astype(np.int64)[np.newaxis, :],
+                "attention_mask": np.array(input_mask).astype(np.int64)[np.newaxis, :],
+                "token_type_ids": np.array(segment_ids).astype(np.int64)[np.newaxis, :],
+            }
+        else:
+            fd = {
+                "input_ids": np.array(input_ids).astype(np.int64)[np.newaxis, :],
+                "input_mask": np.array(input_mask).astype(np.int64)[np.newaxis, :],
+                "segment_ids": np.array(segment_ids).astype(np.int64)[np.newaxis, :],
+            }
+
+        scores = self.sess.run([o.name for o in self.sess.get_outputs()], fd)
+        output = np.stack(scores, axis=-1)[0]
+
+        if self.network == "sut":
+            return output.tolist()
+
+        response_array = array.array("B", output.tobytes())
+        bi = response_array.buffer_info()
+        response = lg.QuerySampleResponse(query_id, bi[0], bi[1])
+        lg.QuerySamplesComplete([response])
+
+    def flush_queries(self):
+        pass
+
+    def __del__(self):
+        if self.profile:
+            print(
+                "ONNX runtime profile dumped to: '{}'".format(
+                    self.sess.end_profiling())
+            )
+        print("Finished destroying SUT.")
+
+
+def get_onnxruntime_sut(args):
+    return BERT_ONNXRuntime_SUT(args)

language/bert/pytorch_SUT.py

+# coding=utf-8
+# Copyright 2021 Arm Limited and affiliates.
+# Copyright (c) 2020 NVIDIA CORPORATION. All rights reserved.
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import array
+import json
+import os
+import sys
+
+sys.path.insert(
+    0,
+    os.path.join(
+        os.getcwd(), "DeepLearningExamples", "PyTorch", "LanguageModeling", "BERT"
+    ),
+)
+sys.path.insert(0, os.getcwd())
+
+try:
+    from squad_QSL import get_squad_QSL
+    from transformers import BertConfig, BertForQuestionAnswering
+    import transformers
+    import torch
+    import numpy as np
+    import mlperf_loadgen as lg
+except ImportError:
+    raise Exception("Error importing local modules")
+
+
+class BERT_PyTorch_SUT:
+    def __init__(self, args):
+        print("Loading BERT configs...")
+        with open("bert_config.json") as f:
+            config_json = json.load(f)
+
+        config = BertConfig(
+            attention_probs_dropout_prob=config_json["attention_probs_dropout_prob"],
+            hidden_act=config_json["hidden_act"],
+            hidden_dropout_prob=config_json["hidden_dropout_prob"],
+            hidden_size=config_json["hidden_size"],
+            initializer_range=config_json["initializer_range"],
+            intermediate_size=config_json["intermediate_size"],
+            max_position_embeddings=config_json["max_position_embeddings"],
+            num_attention_heads=config_json["num_attention_heads"],
+            num_hidden_layers=config_json["num_hidden_layers"],
+            type_vocab_size=config_json["type_vocab_size"],
+            vocab_size=config_json["vocab_size"],
+        )
+
+        self.network = args.network
+        self.dev = (
+            torch.device("cuda:0") if torch.cuda.is_available(
+            ) else torch.device("cpu")
+        )
+        self.version = transformers.__version__
+
+        print("Loading PyTorch model...")
+        self.model = BertForQuestionAnswering(config)
+        self.model.to(self.dev)
+        self.model.eval()
+        model_file = os.environ.get(
+            "ML_MODEL_FILE_WITH_PATH",
+            "build/data/bert_tf_v1_1_large_fp32_384_v2/model.pytorch",
+        )
+        self.model.load_state_dict(torch.load(model_file), strict=False)
+
+        print("Constructing SUT...")
+        self.sut = lg.ConstructSUT(self.issue_queries, self.flush_queries)
+        print("Finished constructing SUT.")
+
+        self.qsl = get_squad_QSL(args.max_examples)
+
+    def issue_queries(self, query_samples):
+        for i in range(len(query_samples)):
+            eval_features = self.qsl.get_features(query_samples[i].index)
+            self.process_sample(eval_features, query_samples[i].id)
+
+    def process_sample(self, sample_input, query_id=None):
+
+        if self.network == "sut":
+            input_ids = sample_input["input_ids"]
+            input_mask = sample_input["input_mask"]
+            segment_ids = sample_input["segment_ids"]
+        else:
+            input_ids = sample_input.input_ids
+            input_mask = sample_input.input_mask
+            segment_ids = sample_input.segment_ids
+
+        with torch.no_grad():
+            model_output = self.model.forward(
+                input_ids=torch.LongTensor(
+                    input_ids).unsqueeze(0).to(self.dev),
+                attention_mask=torch.LongTensor(
+                    input_mask).unsqueeze(0).to(self.dev),
+                token_type_ids=torch.LongTensor(
+                    segment_ids).unsqueeze(0).to(self.dev),
+            )
+            if self.version >= "4.0.0":
+                start_scores = model_output.start_logits
+                end_scores = model_output.end_logits
+            else:
+                start_scores, end_scores = model_output
+            output = (
+                torch.stack([start_scores, end_scores], axis=-1)
+                .squeeze(0)
+                .cpu()
+                .numpy()
+            )
+
+            if self.network == "sut":
+                return output.tolist()
+
+            response_array = array.array("B", output.tobytes())
+            bi = response_array.buffer_info()
+            response = lg.QuerySampleResponse(query_id, bi[0], bi[1])
+            lg.QuerySamplesComplete([response])
+
+    def flush_queries(self):
+        pass
+
+    def __del__(self):
+        print("Finished destroying SUT.")
+
+
+def get_pytorch_sut(args):
+    return BERT_PyTorch_SUT(args)

language/bert/ray_SUT.py

+# coding=utf-8
+# Copyright 2021 Arm Limited and affiliates.
+# Copyright (c) 2020 NVIDIA CORPORATION. All rights reserved.
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import array
+import json
+import os
+import sys
+
+sys.path.insert(
+    0,
+    os.path.join(
+        os.getcwd(), "DeepLearningExamples", "PyTorch", "LanguageModeling", "BERT"
+    ),
+)
+sys.path.insert(0, os.getcwd())
+
+try:
+    from ray.util.actor_pool import ActorPool
+    import ray
+    from squad_QSL import get_squad_QSL
+    from transformers import BertConfig, BertForQuestionAnswering
+    import transformers
+    import torch_tensorrt
+    import torch
+    import numpy as np
+    import mlperf_loadgen as lg
+except ImportError:
+    raise Exception("Error importing local modules")
+
+# Adjustable Parameters
+# Note. num_samples (called "test_query_count" in CM) must be a multiple
+# of batch_size
+BATCH_SIZE = 16
+
+
+@ray.remote(num_cpus=1, num_gpus=1)
+class TorchPredictor:
+    def __init__(self, config_json, model_file, batch_size):
+        print("init", os.getpid(), torch.cuda.device_count())
+        self.pid = os.getpid()
+        self.dev_cnt = torch.cuda.device_count()
+
+        config = BertConfig(
+            attention_probs_dropout_prob=config_json["attention_probs_dropout_prob"],
+            hidden_act=config_json["hidden_act"],
+            hidden_dropout_prob=config_json["hidden_dropout_prob"],
+            hidden_size=config_json["hidden_size"],
+            initializer_range=config_json["initializer_range"],
+            intermediate_size=config_json["intermediate_size"],
+            max_position_embeddings=config_json["max_position_embeddings"],
+            num_attention_heads=config_json["num_attention_heads"],
+            num_hidden_layers=config_json["num_hidden_layers"],
+            type_vocab_size=config_json["type_vocab_size"],
+            vocab_size=config_json["vocab_size"],
+        )
+
+        self.dev = torch.device("cuda")
+        self.model = BertForQuestionAnswering(config)
+        self.model.to(self.dev)
+        self.model.eval()
+        self.model.load_state_dict(torch.load(model_file), strict=False)
+        # tensor rt
+        batch_input_ids = torch.LongTensor(
+            np.zeros(
+                (batch_size, 384))).to(
+            self.dev)
+        traced_mlm_model = torch.jit.trace(
+            self.model,
+            [batch_input_ids, batch_input_ids, batch_input_ids],
+            strict=False,
+        )
+        self.trt_model = torch_tensorrt.compile(
+            traced_mlm_model,
+            inputs=[
+                torch_tensorrt.Input(
+                    shape=[
+                        batch_size,
+                        384],
+                    dtype=torch.int32),
+                torch_tensorrt.Input(
+                    shape=[
+                        batch_size,
+                        384],
+                    dtype=torch.int32),
+                torch_tensorrt.Input(
+                    shape=[
+                        batch_size,
+                        384],
+                    dtype=torch.int32),
+            ],
+            enabled_precisions={torch.float32, torch.float16},
+            workspace_size=2000000000,
+            truncate_long_and_double=True,
+        )
+
+        print("done loading")
+
+    # Logic for inference on 1 batch of data.
+    def forward(self, batch):
+        input_ids = torch.from_numpy(batch["input_ids"]).to(self.dev)
+        attention_mask = torch.from_numpy(batch["attention_mask"]).to(self.dev)
+        token_type_ids = torch.from_numpy(batch["token_type_ids"]).to(self.dev)
+        with torch.inference_mode():
+            # pytorch
+            # model_output = self.model.forward(input_ids=input_ids,
+            #     attention_mask=attention_mask,
+            #     token_type_ids=token_type_ids)
+            # start_scores = model_output.start_logits
+            # end_scores = model_output.end_logits
+
+            # tensor rt
+            trt_output = self.trt_model(
+                input_ids, attention_mask, token_type_ids)
+            start_scores = trt_output["start_logits"]
+            end_scores = trt_output["end_logits"]
+
+            batch_ret = torch.stack(
+                [start_scores, end_scores], axis=-1).cpu().numpy()
+            return {"output": batch_ret}
+
+    def ready(self):
+        pass
+
+
+class BERT_Ray_SUT:
+    def __init__(self, args):
+        with open("bert_config.json") as f:
+            config_json = json.load(f)
+        model_file = os.environ.get(
+            "ML_MODEL_FILE_WITH_PATH",
+            "build/data/bert_tf_v1_1_large_fp32_384_v2/model.pytorch",
+        )
+
+        print("Constructing SUT...")
+        self.sut = lg.ConstructSUT(self.issue_queries, self.flush_queries)
+        print("Finished constructing SUT.")
+        self.qsl = get_squad_QSL(args.max_examples)
+
+        try:
+            ray.init(address="auto")
+        except BaseException:
+            print("WARN: Cannot connect to existing Ray cluster.")
+            print("We are going to start a new RAY cluster, but pay attention that")
+            print("the cluster contains only one node.")
+            print(
+                "If you want to use multiple nodes, please start the cluster manually via:"
+            )
+            print("\tOn the head node, run `ray start --head`")
+            print("\tOn other nodes, run `ray start --address=<head node IP>:6379`")
+            ray.init()
+
+        self.batch_size = BATCH_SIZE
+        resources = ray.cluster_resources()
+        num_gpus = int(resources.get("GPU", 0))
+
+        print(f"The cluster has {num_gpus} GPUs.")
+
+        self.actor_list = [
+            TorchPredictor.remote(config_json, model_file, self.batch_size)
+            for _ in range(num_gpus)
+        ]
+        self.pool = ActorPool(self.actor_list)
+
+        samples = []
+        for i in range(self.qsl.count):
+            sample = {}
+            eval_features = self.qsl.get_features(i)
+            sample["input_ids"] = np.array(
+                eval_features.input_ids).astype(
+                np.int32)
+            sample["attention_mask"] = np.array(eval_features.input_mask).astype(
+                np.int32
+            )
+            sample["token_type_ids"] = np.array(eval_features.segment_ids).astype(
+                np.int32
+            )
+            samples.append(sample)
+        self.samples = samples
+
+        print("Waiting Actors init")
+        for actor in self.actor_list:
+            ray.get(actor.ready.remote())
+        print("BERT_Ray_SUT construct complete")
+
+    def issue_queries(self, query_samples):
+        if len(query_samples) % self.batch_size != 0:
+            print("ERROR: batch size must be a multiple of the number of samples")
+            sys.exit(1)
+
+        batch_samples = []
+        i = 0
+        while i < len(query_samples):
+            batch_sample = {
+                "input_ids": np.array(
+                    [
+                        self.samples[query_sample.index]["input_ids"]
+                        for query_sample in query_samples[i: i + self.batch_size]
+                    ]
+                ),
+                "attention_mask": np.array(
+                    [
+                        self.samples[query_sample.index]["attention_mask"]
+                        for query_sample in query_samples[i: i + self.batch_size]
+                    ]
+                ),
+                "token_type_ids": np.array(
+                    [
+                        self.samples[query_sample.index]["token_type_ids"]
+                        for query_sample in query_samples[i: i + self.batch_size]
+                    ]
+                ),
+            }
+            batch_samples.append(batch_sample)
+            i = i + self.batch_size
+
+        # print("samples len", len(batch_samples))
+        batch_inference_results = list(
+            self.pool.map_unordered(
+                lambda a, v: a.forward.remote(v), batch_samples)
+        )
+
+        cur_query_index = 0
+        for batch_inference_result in batch_inference_results:
+            batch_inference_result = batch_inference_result["output"]
+            for inference_result in batch_inference_result:
+                response_array = array.array("B", inference_result.tobytes())
+                bi = response_array.buffer_info()
+                response = lg.QuerySampleResponse(
+                    query_samples[cur_query_index].id, bi[0], bi[1]
+                )
+                lg.QuerySamplesComplete([response])
+                cur_query_index += 1
+
+    def flush_queries(self):
+        pass
+
+    def __del__(self):
+        print("Finished destroying SUT.")
+
+
+def get_ray_sut(args):
+    return BERT_Ray_SUT(args)

language/bert/run.py

+# coding=utf-8
+# Copyright 2021 Arm Limited and affiliates.
+# Copyright (c) 2020 NVIDIA CORPORATION. All rights reserved.
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from absl import flags
+from absl import app
+import subprocess
+import mlperf_loadgen as lg
+import argparse
+import os
+import sys
+
+sys.path.insert(0, os.getcwd())
+sys.path.insert(0, os.path.join(os.getcwd(), "..", "..", "lon"))
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--backend",
+        choices=["tf", "pytorch", "onnxruntime", "tf_estimator", "ray"],
+        default="tf",
+        help="Backend",
+    )
+    parser.add_argument(
+        "--scenario",
+        choices=["SingleStream", "Offline", "Server", "MultiStream"],
+        default="Offline",
+        help="Scenario",
+    )
+    parser.add_argument(
+        "--accuracy",
+        action="store_true",
+        help="enable accuracy pass")
+    parser.add_argument(
+        "--quantized",
+        action="store_true",
+        help="use quantized model (only valid for onnxruntime backend)",
+    )
+    parser.add_argument(
+        "--profile",
+        action="store_true",
+        help="enable profiling (only valid for onnxruntime backend)",
+    )
+    parser.add_argument(
+        "--user_conf",
+        default="user.conf",
+        help="user config for user LoadGen settings such as target QPS",
+    )
+    parser.add_argument(
+        "--audit_conf",
+        default="audit.conf",
+        help="audit config for LoadGen settings during compliance runs",
+    )
+    parser.add_argument(
+        "--max_examples",
+        type=int,
+        help="Maximum number of examples to consider (not limited by default)",
+    )
+    parser.add_argument(
+        "--network",
+        choices=["sut", "lon", None],
+        default=None,
+        help="Loadgen network mode",
+    )
+    parser.add_argument("--node", type=str, default="")
+    parser.add_argument("--port", type=int, default=8000)
+    parser.add_argument(
+        "--sut_server",
+        nargs="*",
+        default=["http://localhost:8000"],
+        help="Address of the server(s) under test.",
+    )
+
+    args = parser.parse_args()
+    return args
+
+
+scenario_map = {
+    "SingleStream": lg.TestScenario.SingleStream,
+    "Offline": lg.TestScenario.Offline,
+    "Server": lg.TestScenario.Server,
+    "MultiStream": lg.TestScenario.MultiStream,
+}
+
+
+def main():
+    args = get_args()
+
+    sut = None
+
+    if not args.network or args.network == "sut":
+        if args.backend == "pytorch":
+            assert (
+                not args.quantized
+            ), "Quantized model is only supported by onnxruntime backend!"
+            assert (
+                not args.profile
+            ), "Profiling is only supported by onnxruntime backend!"
+            from pytorch_SUT import get_pytorch_sut
+
+            sut = get_pytorch_sut(args)
+        elif args.backend == "tf":
+            assert (
+                not args.quantized
+            ), "Quantized model is only supported by onnxruntime backend!"
+            assert (
+                not args.profile
+            ), "Profiling is only supported by onnxruntime backend!"
+            from tf_SUT import get_tf_sut
+
+            sut = get_tf_sut(args)
+        elif args.backend == "tf_estimator":
+            assert (
+                not args.quantized
+            ), "Quantized model is only supported by onnxruntime backend!"
+            assert (
+                not args.profile
+            ), "Profiling is only supported by onnxruntime backend!"
+            from tf_estimator_SUT import get_tf_estimator_sut
+
+            sut = get_tf_estimator_sut()
+        elif args.backend == "onnxruntime":
+            from onnxruntime_SUT import get_onnxruntime_sut
+
+            sut = get_onnxruntime_sut(args)
+        elif args.backend == "ray":
+            assert (
+                not args.quantized
+            ), "Quantized model is only supported by onnxruntime backend!"
+            assert (
+                not args.profile
+            ), "Profiling is only supported by onnxruntime backend!"
+            from ray_SUT import get_ray_sut
+
+            sut = get_ray_sut(args)
+        else:
+            raise ValueError("Unknown backend: {:}".format(args.backend))
+
+    settings = lg.TestSettings()
+    settings.scenario = scenario_map[args.scenario]
+    # mlperf.conf is automatically loaded by the loadgen
+    # settings.FromConfig(args.mlperf_conf, "bert", args.scenario)
+    settings.FromConfig(args.user_conf, "bert", args.scenario)
+
+    if args.accuracy:
+        settings.mode = lg.TestMode.AccuracyOnly
+    else:
+        settings.mode = lg.TestMode.PerformanceOnly
+    log_path = os.environ.get("LOG_PATH")
+    if not log_path:
+        log_path = "build/logs"
+    if not os.path.exists(log_path):
+        os.makedirs(log_path)
+    log_output_settings = lg.LogOutputSettings()
+    log_output_settings.outdir = log_path
+    log_output_settings.copy_summary_to_stdout = True
+    log_settings = lg.LogSettings()
+    log_settings.log_output = log_output_settings
+    log_settings.enable_trace = True
+
+    if args.network == "lon":
+        from network_LON import app, set_args, main as app_main
+
+        set_args(
+            args,
+            settings,
+            log_settings,
+            args.audit_conf,
+            args.sut_server,
+            args.backend,
+            args.max_examples,
+        )
+        app.run(app_main)
+
+    elif args.network == "sut":
+        from network_SUT import app, node, set_backend
+
+        node = args.node
+        set_backend(sut)
+        app.run(debug=False, port=args.port, host="0.0.0.0")
+
+    else:
+        print("Running LoadGen test...")
+        lg.StartTestWithLogSettings(
+            sut.sut, sut.qsl.qsl, settings, log_settings, args.audit_conf
+        )
+        if args.accuracy and not os.environ.get("SKIP_VERIFY_ACCURACY"):
+            cmd = "python3 {:}/accuracy-squad.py {}".format(
+                os.path.dirname(os.path.abspath(__file__)),
+                (
+                    "--max_examples {}".format(args.max_examples)
+                    if args.max_examples
+                    else ""
+                ),
+            )
+            subprocess.check_call(cmd, shell=True)
+
+    print("Done!")
+
+    if sut:
+        print("Destroying SUT...")
+        lg.DestroySUT(sut.sut)
+
+        print("Destroying QSL...")
+        lg.DestroyQSL(sut.qsl.qsl)
+
+
+if __name__ == "__main__":
+    main()