Allow bert_perf_test.py to load/save tuning results (#15096)

onnxruntime/python/tools/transformers/bert_perf_test.py

@@ -14,6 +14,7 @@
 
 import argparse
 import csv
+import json
 import multiprocessing
 import os
 import random
@@ -22,6 +23,7 @@ import timeit
 from dataclasses import dataclass
 from datetime import datetime
 from pathlib import Path
+from typing import Optional
 
 import numpy as np
 import psutil
@@ -51,9 +53,19 @@ class ModelSetting:
     segment_ids_name: str
     input_mask_name: str
     opt_level: int
+    input_tuning_results: Optional[str]
+    output_tuning_results: Optional[str]
 
 
-def create_session(model_path, use_gpu, provider, intra_op_num_threads, graph_optimization_level=None, log_severity=2):
+def create_session(
+    model_path,
+    use_gpu,
+    provider,
+    intra_op_num_threads,
+    graph_optimization_level=None,
+    log_severity=2,
+    tuning_results_path=None,
+):
     import onnxruntime
 
     onnxruntime.set_default_logger_severity(log_severity)
@@ -127,6 +139,10 @@ def create_session(model_path, use_gpu, provider, intra_op_num_threads, graph_op
     else:
         assert "CPUExecutionProvider" in session.get_providers()
 
+    if tuning_results_path is not None:
+        with open(tuning_results_path) as f:
+            session.set_tuning_results(json.load(f))
+
     return session
 
 
@@ -228,6 +244,7 @@ def run_one_test(model_setting, test_setting, perf_results, all_inputs, intra_op
         intra_op_num_threads,
         model_setting.opt_level,
         log_severity=test_setting.log_severity,
+        tuning_results_path=model_setting.input_tuning_results,
     )
     output_names = [output.name for output in session.get_outputs()]
 
@@ -275,6 +292,18 @@ def run_one_test(model_setting, test_setting, perf_results, all_inputs, intra_op
         "Average latency = {} ms, Throughput = {} QPS".format(format(average_latency, ".2f"), format(throughput, ".2f"))
     )
 
+    if model_setting.output_tuning_results:
+        output_path = os.path.abspath(model_setting.output_tuning_results)
+        if os.path.exists(output_path):
+            old_output_path = output_path
+            output_path = f"""{output_path.rsplit(".json", 1)[0]}.{datetime.now().timestamp()}.json"""
+            print("WARNING:", old_output_path, "exists, will write to", output_path, "instead.")
+
+        trs = session.get_tuning_results()
+        with open(output_path, "w") as f:
+            json.dump(trs, f)
+        print("Tuning results is saved to", output_path)
+
 
 def launch_test(model_setting, test_setting, perf_results, all_inputs, intra_op_num_threads):
     process = multiprocessing.Process(
@@ -459,6 +488,19 @@ def parse_arguments():
         help="input name for attention mask",
     )
 
+    parser.add_argument(
+        "--input_tuning_results",
+        default=None,
+        type=str,
+        help="tuning results (json) to be loaded before benchmark",
+    )
+    parser.add_argument(
+        "--output_tuning_results",
+        default=None,
+        type=str,
+        help="tuning results (json) to be saved after benchmark",
+    )
+
     args = parser.parse_args()
     return args
 
@@ -482,6 +524,8 @@ def main():
         args.segment_ids_name,
         args.input_mask_name,
         args.opt_level,
+        args.input_tuning_results,
+        args.output_tuning_results,
     )
 
     for batch_size in batch_size_set: