intel
diff --git a/‎CODEOWNERS
Lines changed: 1 addition & 0 deletions b/‎CODEOWNERS
Lines changed: 1 addition & 0 deletions
diff --git a/‎README.md
Lines changed: 1 addition & 1 deletion b/‎README.md
Lines changed: 1 addition & 1 deletion
diff --git a/‎benchmarks/common/base_benchmark_util.py
Lines changed: 18 additions & 3 deletions b/‎benchmarks/common/base_benchmark_util.py
Lines changed: 18 additions & 3 deletions
diff --git a/‎benchmarks/common/tensorflow/start.sh
Lines changed: 93 additions & 6 deletions b/‎benchmarks/common/tensorflow/start.sh
Lines changed: 93 additions & 6 deletions
diff --git a/‎benchmarks/image_recognition/tensorflow/resnet50v1_5/inference/bfloat16/model_init.py
Lines changed: 28 additions & 13 deletions b/‎benchmarks/image_recognition/tensorflow/resnet50v1_5/inference/bfloat16/model_init.py
Lines changed: 28 additions & 13 deletions
diff --git a/‎benchmarks/image_recognition/tensorflow/resnet50v1_5/inference/fp16/__init__.py
Lines changed: 2 additions & 0 deletions b/‎benchmarks/image_recognition/tensorflow/resnet50v1_5/inference/fp16/__init__.py
Lines changed: 2 additions & 0 deletions
diff --git a/‎benchmarks/image_recognition/tensorflow/resnet50v1_5/inference/fp16/model_init.py
Lines changed: 29 additions & 15 deletions b/‎benchmarks/image_recognition/tensorflow/resnet50v1_5/inference/fp16/model_init.py
Lines changed: 29 additions & 15 deletions
@@ -8,6 +8,7 @@ datasets @ashahba @claynerobison @dzungductran
 docs @claynerobison @mhbuehler
 k8s  @ashahba @dzungductran
 models @ashraf-bhuiyan @riverliuintel
+models @riverliuintel
 models/**/pytorch/ @leslie-fang-intel @jiayisunx @zhuhaozhe
 quickstart mahathi.vatsal.salopanthula@intel.com
 quickstart/**/pytorch/ @leslie-fang-intel @jiayisunx @zhuhaozhe
 
@@ -1,6 +1,6 @@
 # Model Zoo for Intel® Architecture
 
-This repository contains **links to pre-trained models, sample scripts, best practices, and step-by-step tutorials** for many popular open-source machine learning models optimized by Intel to run on Intel® Xeon® Scalable processors.
+This repository contains **links to pre-trained models, sample scripts, best practices, and step-by-step tutorials** for many popular open-source machine learning models optimized by Intel to run on Intel® Xeon® Scalable processors and Intel® Data Center GPUs.
 
 Model packages and containers for running the Model Zoo's workloads can be found at the [Intel® Developer Catalog](https://software.intel.com/containers).
 
 
@@ -1,7 +1,7 @@
 #
 # -*- coding: utf-8 -*-
 #
-# Copyright (c) 2023 Intel Corporation
+# Copyright (c) 2018-2023 Intel Corporation
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -281,6 +281,12 @@ def _define_args(self):
             help="Additional command line arguments (prefix flag start with"
                  " '--').")
 
+        # Check if GPU is enabled.
+        self._common_arg_parser.add_argument(
+            "--gpu",
+            help="Run the benchmark script using GPU",
+            dest="gpu", action="store_true")
+
     def _validate_args(self):
         """validate the args and initializes platform_util"""
         # check if socket id is in socket number range
@@ -311,8 +317,9 @@ def _validate_args(self):
                              format(system_num_cores))
 
         if args.output_results and ((args.model_name != "resnet50" and
-                                    args.model_name != "resnet50v1_5") or args.precision != "fp32"):
-            raise ValueError("--output-results is currently only supported for resnet50 FP32 inference.")
+                                    args.model_name != "resnet50v1_5") or
+                                    (args.precision != "fp32" and args.precision != "fp16")):
+            raise ValueError("--output-results is currently only supported for resnet50 FP32 or FP16 inference.")
         elif args.output_results and (args.mode != "inference" or not args.data_location):
             raise ValueError("--output-results can only be used when running inference with a dataset.")
 
@@ -355,6 +362,14 @@ def _validate_args(self):
                       "This is less than the number of cores per socket on the system ({})".
                       format(args.socket_id, cpuset_len_for_socket, self._platform_util.num_cores_per_socket))
 
+        if args.gpu:
+            if args.socket_id != -1:
+                raise ValueError("--socket-id cannot be used with --gpu parameter.")
+            if args.num_intra_threads is not None:
+                raise ValueError("--num-intra-threads cannot be used with --gpu parameter.")
+            if args.num_inter_threads is not None:
+                raise ValueError("--num-inter-threads cannot be used with --gpu parameter.")
+
     def initialize_model(self, args, unknown_args):
         """Create model initializer for the specified model"""
         model_initializer = None
 
@@ -1,6 +1,6 @@
 #!/usr/bin/env bash
 #
-# Copyright (c) 2023 Intel Corporation
+# Copyright (c) 2018-2023 Intel Corporation
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -54,7 +54,26 @@ echo "    NUMA_CORES_PER_INSTANCE: ${NUMA_CORES_PER_INSTANCE}"
 echo "    PYTHON_EXE: ${PYTHON_EXE}"
 echo "    PYTHONPATH: ${PYTHONPATH}"
 echo "    DRY_RUN: ${DRY_RUN}"
-
+echo "    GPU: ${GPU}"
+
+#  Enable GPU Flag
+gpu_arg=""
+is_model_gpu_supported="False"
+if [ ${GPU} == "True" ]; then
+  gpu_arg="--gpu"
+  # Environment variables for GPU
+  export RenderCompressedBuffersEnabled=0
+  export CreateMultipleSubDevices=1
+  export ForceLocalMemoryAccessMode=1
+  export SYCL_PI_LEVEL_ZERO_BATCH_SIZE=1
+else
+  unset RenderCompressedBuffersEnabled
+  unset CreateMultipleSubDevices
+  unset ForceLocalMemoryAccessMode
+  unset ForceNonSystemMemoryPlacement
+  unset TF_ENABLE_LAYOUT_OPT
+  unset SYCL_PI_LEVEL_ZERO_BATCH_SIZE
+fi
 #  inference & training is supported right now
 if [ ${MODE} != "inference" ] && [ ${MODE} != "training" ]; then
   echo "${MODE} mode for ${MODEL_NAME} is not supported"
@@ -326,6 +345,10 @@ export PYTHONPATH=${PYTHONPATH}:${MOUNT_INTELAI_MODELS_COMMON_SOURCE}:${MOUNT_IN
 
 # Common execution command used by all models
 function run_model() {
+  if [ ${is_model_gpu_supported} == "False"  ] && [ ${GPU} == "True" ]; then
+    echo "Runing ${MODEL_NAME} ${MODE} with precision ${PRECISION} does not support --gpu."
+    exit 1
+  fi
   # Navigate to the main benchmark directory before executing the script,
   # since the scripts use the benchmark/common scripts as well.
   cd ${MOUNT_BENCHMARK}
@@ -390,7 +413,8 @@ ${benchmark_only_arg} \
 ${output_results_arg} \
 ${weight_sharing_arg} \
 ${synthetic_data_arg} \
-${verbose_arg}"
+${verbose_arg} \
+${gpu_arg}"
 
 if [ ${MOUNT_EXTERNAL_MODELS_SOURCE} != "None" ]; then
   CMD="${CMD} --model-source-dir=${MOUNT_EXTERNAL_MODELS_SOURCE}"
@@ -978,6 +1002,7 @@ function resnet101_inceptionv3() {
 # ResNet50  model
 function resnet50() {
     export PYTHONPATH=${PYTHONPATH}:$(pwd):${MOUNT_BENCHMARK}
+    is_model_gpu_supported="True"
 
     # For accuracy, dataset location is required.
     if [ "${DATASET_LOCATION_VOL}" == "None" ] && [ ${ACCURACY_ONLY} == "True" ]; then
@@ -1062,6 +1087,7 @@ function rfcn() {
 
 # SSD-MobileNet model
 function ssd_mobilenet() {
+  is_model_gpu_supported="True"
   if [ ${PRECISION} == "fp32" ] || [ ${PRECISION} == "bfloat16" ]; then
     if [ ${BATCH_SIZE} != "-1" ]; then
       echo "Warning: SSD-MobileNet FP32 inference script does not use the batch_size arg"
@@ -1404,7 +1430,21 @@ function wavenet() {
 
 # BERT base
 function bert_base() {
-  if [ ${PRECISION} == "fp32" ]  || [ $PRECISION == "bfloat16" ]; then
+  if [ ${GPU} == "True" ]; then
+    if [ ${MODE} == "inference" ]; then
+      echo "PRECISION=${PRECISION} on GPU not supported for ${MODEL_NAME} ${MODE} in this repo."
+      exit 1
+    elif [ ${MODE} == "training" ]; then
+      if [ ${PRECISION} != "fp32" ] && [ ${PRECISION} != "bfloat16" ]; then
+        echo "PRECISION=${PRECISION} on GPU not supported for ${MODEL_NAME} ${MODE} in this repo."
+        exit 1
+      fi
+    fi
+    is_model_gpu_supported="True"
+    export PYTHONPATH=${PYTHONPATH}:${MOUNT_EXTERNAL_MODELS_SOURCE}
+    bert_options
+    CMD=${CMD} run_model
+  elif [ ${PRECISION} == "fp32" ]  || [ $PRECISION == "bfloat16" ]; then
     export PYTHONPATH=${PYTHONPATH}:${MOUNT_EXTERNAL_MODELS_SOURCE}
     bert_options
     CMD=${CMD} run_model
@@ -1416,11 +1456,58 @@ function bert_base() {
 
 # BERT Large model
 function bert_large() {
-    # Change if to support fp32
-    if [ ${PRECISION} == "fp32" ]  || [ $PRECISION == "int8" ] || [ $PRECISION == "bfloat16" ] || [ $PRECISION == "fp16" ]; then
+    export PYTHONPATH=${PYTHONPATH}:${MOUNT_BENCHMARK}
+    if [ ${GPU} == "True" ]; then
+      if [ ${MODE} == "inference" ]; then
+        if [ ${PRECISION} != "fp32" ] && [ ${PRECISION} != "fp16" ] && [ ${PRECISION} != "bfloat16" ]; then
+          echo "PRECISION=${PRECISION} on GPU not supported for ${MODEL_NAME} ${MODE} in this repo."
+          exit 1
+        fi
+      elif [ ${MODE} == "training" ]; then
+        if [ ${PRECISION} != "fp32" ] && [ ${PRECISION} != "bfloat16" ]; then
+          echo "PRECISION=${PRECISION} on GPU not supported for ${MODEL_NAME} ${MODE} in this repo."
+          exit 1
+        fi
+      fi
+      is_model_gpu_supported="True"
       export PYTHONPATH=${PYTHONPATH}:${MOUNT_EXTERNAL_MODELS_SOURCE}
       bert_options
       CMD=${CMD} run_model
+    else
+      if [ ${PRECISION} == "fp32" ]  || [ $PRECISION == "int8" ] || [ $PRECISION == "bfloat16" ] || [ $PRECISION == "fp16" ]; then
+        export PYTHONPATH=${PYTHONPATH}:${MOUNT_EXTERNAL_MODELS_SOURCE}
+        bert_options
+        CMD=${CMD} run_model
+      else
+        echo "PRECISION=${PRECISION} not supported for ${MODEL_NAME} in this repo."
+        exit 1
+      fi
+    fi
+}
+
+# distilBERT base model
+function distilbert_base() {
+    if [ ${PRECISION} == "fp32" ] || [ ${PRECISION} == "bfloat16" ]|| [ ${PRECISION} == "int8" ]; then
+      export PYTHONPATH=${PYTHONPATH}:${MOUNT_EXTERNAL_MODELS_SOURCE}
+      CMD="${CMD} $(add_arg "--warmup-steps" ${WARMUP_STEPS})"
+      CMD="${CMD} $(add_arg "--steps" ${STEPS})"
+
+      if [ ${NUM_INTER_THREADS} != "None" ]; then
+        CMD="${CMD} $(add_arg "--num-inter-threads" ${NUM_INTER_THREADS})"
+      fi
+
+      if [ ${NUM_INTRA_THREADS} != "None" ]; then
+        CMD="${CMD} $(add_arg "--num-intra-threads" ${NUM_INTRA_THREADS})"
+      fi
+
+      if [ -z ${STEPS} ]; then
+        CMD="${CMD} $(add_arg "--steps" ${STEPS})"
+      fi
+
+      if [ -z $MAX_SEQ_LENGTH ]; then
+        CMD="${CMD} $(add_arg "--max-seq-length" ${MAX_SEQ_LENGTH})"
+      fi
+      CMD=${CMD} run_model
     else
       echo "PRECISION=${PRECISION} not supported for ${MODEL_NAME} in this repo."
       exit 1
 
@@ -65,33 +65,48 @@ def __init__(self, args, custom_args=[], platform_util=None):
         config_file_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "config.json")
         self.set_kmp_vars(config_file_path, kmp_blocktime=str(self.args.kmp_blocktime))
 
-        set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
+        if not self.args.gpu:
+            set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
 
         # If weight-sharing flag is ON, then use the weight-sharing script.
         if self.args.weight_sharing and not self.args.accuracy_only:
             benchmark_script = os.path.join(
                 self.args.intelai_models, self.args.mode,
                 "eval_image_classifier_inference_weight_sharing.py")
         else:
-            benchmark_script = os.path.join(
-                self.args.intelai_models, self.args.mode,
-                "eval_image_classifier_inference.py")
+            if self.args.gpu:
+                benchmark_script = os.path.join(
+                    self.args.intelai_models, self.args.mode, self.args.precision,
+                    "eval_image_classifier_inference.py")
+            else:
+                benchmark_script = os.path.join(
+                    self.args.intelai_models, self.args.mode,
+                    "eval_image_classifier_inference.py")
 
         self.benchmark_command = self.get_command_prefix(args.socket_id) + \
             self.python_exe + " " + benchmark_script
 
         num_cores = self.platform_util.num_cores_per_socket if self.args.num_cores == -1 \
             else self.args.num_cores
 
-        self.benchmark_command = \
-            self.benchmark_command + \
-            " --input-graph=" + self.args.input_graph + \
-            " --num-inter-threads=" + str(self.args.num_inter_threads) + \
-            " --num-intra-threads=" + str(self.args.num_intra_threads) + \
-            " --num-cores=" + str(num_cores) + \
-            " --batch-size=" + str(self.args.batch_size) + \
-            " --warmup-steps=" + str(self.args.warmup_steps) + \
-            " --steps=" + str(self.args.steps)
+        if self.args.gpu:
+            self.benchmark_command = \
+                self.benchmark_command + \
+                " --input-graph=" + self.args.input_graph + \
+                " --num-cores=" + str(num_cores) + \
+                " --batch-size=" + str(self.args.batch_size) + \
+                " --warmup-steps=" + str(self.args.warmup_steps) + \
+                " --steps=" + str(self.args.steps)
+        else:
+            self.benchmark_command = \
+                self.benchmark_command + \
+                " --input-graph=" + self.args.input_graph + \
+                " --num-inter-threads=" + str(self.args.num_inter_threads) + \
+                " --num-intra-threads=" + str(self.args.num_intra_threads) + \
+                " --num-cores=" + str(num_cores) + \
+                " --batch-size=" + str(self.args.batch_size) + \
+                " --warmup-steps=" + str(self.args.warmup_steps) + \
+                " --steps=" + str(self.args.steps)
 
         if self.args.data_num_inter_threads:
             self.benchmark_command += " --data-num-inter-threads=" + str(self.args.data_num_inter_threads)
 
@@ -15,3 +15,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 #
+
+#
@@ -65,34 +65,48 @@ def __init__(self, args, custom_args=[], platform_util=None):
         config_file_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "config.json")
         self.set_kmp_vars(config_file_path, kmp_blocktime=str(self.args.kmp_blocktime))
 
-        set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
-
         # If weight-sharing flag is ON, then use the weight-sharing script.
         if self.args.weight_sharing and not self.args.accuracy_only:
+            set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
             benchmark_script = os.path.join(
                 self.args.intelai_models, self.args.mode,
                 "eval_image_classifier_inference_weight_sharing.py")
         else:
-            benchmark_script = os.path.join(
-                self.args.intelai_models, self.args.mode,
-                "eval_image_classifier_inference.py")
+            if self.args.gpu:
+                benchmark_script = os.path.join(
+                    self.args.intelai_models, self.args.mode, self.args.precision,
+                    "eval_image_classifier_inference.py")
+            else:
+                set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
+                benchmark_script = os.path.join(
+                    self.args.intelai_models, self.args.mode,
+                    "eval_image_classifier_inference.py")
 
         self.benchmark_command = self.get_command_prefix(args.socket_id) + \
             self.python_exe + " " + benchmark_script
 
         num_cores = self.platform_util.num_cores_per_socket if self.args.num_cores == -1 \
             else self.args.num_cores
 
-        self.benchmark_command = \
-            self.benchmark_command + \
-            " --input-graph=" + self.args.input_graph + \
-            " --data-type=" + self.args.precision + \
-            " --num-inter-threads=" + str(self.args.num_inter_threads) + \
-            " --num-intra-threads=" + str(self.args.num_intra_threads) + \
-            " --num-cores=" + str(num_cores) + \
-            " --batch-size=" + str(self.args.batch_size) + \
-            " --warmup-steps=" + str(self.args.warmup_steps) + \
-            " --steps=" + str(self.args.steps)
+        if self.args.gpu:
+            self.benchmark_command = \
+                self.benchmark_command + \
+                " --input-graph=" + self.args.input_graph + \
+                " --num-cores=" + str(num_cores) + \
+                " --batch-size=" + str(self.args.batch_size) + \
+                " --warmup-steps=" + str(self.args.warmup_steps) + \
+                " --steps=" + str(self.args.steps)
+        else:
+            self.benchmark_command = \
+                self.benchmark_command + \
+                " --input-graph=" + self.args.input_graph + \
+                " --data-type=" + self.args.precision + \
+                " --num-inter-threads=" + str(self.args.num_inter_threads) + \
+                " --num-intra-threads=" + str(self.args.num_intra_threads) + \
+                " --num-cores=" + str(num_cores) + \
+                " --batch-size=" + str(self.args.batch_size) + \
+                " --warmup-steps=" + str(self.args.warmup_steps) + \
+                " --steps=" + str(self.args.steps)
 
         if self.args.data_num_inter_threads:
             self.benchmark_command += " --data-num-inter-threads=" + str(self.args.data_num_inter_threads)
Original file line number	Diff line number	Diff line change
`@@ -15,3 +15,5 @@`
`15`	`15`	`# See the License for the specific language governing permissions and`
`16`	`16`	`# limitations under the License.`
`17`	`17`	`#`
	`18`	`+`
	`19`	`+#`