Merge remote-tracking branch 'mlcommons/master'

Author: suachong

VERSION

@@ -1 +1 @@
-4.1.27
+4.1.29

mlperf_logging/result_summarizer/compute_score/README.md

@@ -0,0 +1,30 @@
+# MLPerf compute standalone score
+
+MLPerf compute standalone score
+
+## Usage
+
+To compute the scores of a single benchmark. All the results files are assumed to be in the same folder:
+
+```sh
+python3 -m mlperf_logging.result_summarizer.compute_score --benchmark BENCHMARK \
+    --system SYSTEM_NAME --benchmark_folder BENCHMARK_FOLDER --usage USAGE --ruleset RULESET \
+    [--is_weak_scaling] [--scale] [--has_power]
+```
+
+
+**BENCHMARK:** Name of the benchmark to compute the score such as rgat, llama31_8b, etc.
+**SYSTEM_NAME:** Optional system name.
+**BENCHMARK_FOLDER:** Folder containing all the results files of the benchmark.
+**USAGE:** Either "training" or "hpc",
+**RULESET:** Version of the rules that applies one of "1.0.0", "1.1.0", "2.0.0", "2.1.0", "3.0.0", "3.1.0", "4.0.0", "4.1.0", "5.0.0", "5.1.0".
+**[--is_weak_scaling]:** Is the benchmark weak scaling (only applies to HPC).
+**[--scale]:** Compute the scaling.json file (only if the folder does not contain it already).
+**[--has_power]:** Have the results power measurements .
+
+
+
+## Tested software versions
+Tested and confirmed working using the following software versions:
+
+Python 3.9.18

mlperf_logging/result_summarizer/compute_score/__init__.py

@@ -0,0 +1,188 @@
+from .. import result_summarizer
+from ...rcp_checker import rcp_checker
+from ...compliance_checker.mlp_compliance import usage_choices, rule_choices
+from ...compliance_checker.mlp_parser import parse_file
+from ...benchmark_meta import get_result_file_counts
+import argparse
+import glob
+import json
+import os
+
+
+def get_compute_args():
+    parser = argparse.ArgumentParser(
+        prog="mlperf_logging.result_summarizer.compute_score",
+        description="Compute the score of a single benchmark",
+    )
+    parser.add_argument("--system", type=str, help="System name", default=None)
+    parser.add_argument(
+        "--has_power", action="store_true", help="Compute power score as well"
+    )
+    parser.add_argument(
+        "--benchmark_folder",
+        type=str,
+        help="Folder containing all the result files",
+        required=True,
+    )
+    parser.add_argument(
+        "--usage",
+        type=str,
+        default="training",
+        choices=usage_choices(),
+        help="the usage such as training, hpc, inference_edge, inference_server",
+        required=True,
+    )
+    parser.add_argument(
+        "--ruleset",
+        type=str,
+        choices=rule_choices(),
+        help="the ruleset such as 0.6.0, 0.7.0, or 1.0.0",
+        required=True,
+    )
+    parser.add_argument(
+        "--is_weak_scaling", action="store_true", help="Compute weak scaling score"
+    )
+    parser.add_argument(
+        "--scale", action="store_true", help="Compute the scaling factor"
+    )
+
+    return parser.parse_args()
+
+
+def print_benchmark_info(args, benchmark):
+    print("INFO -------------------------------------------------------")
+    print(f"MLPerf {args.usage}")
+    print(f"Folder: {args.benchmark_folder}")
+    print(f"Version: {args.ruleset}")
+    print(f"System: {args.system}")
+    print(f"Benchmark: {benchmark}")
+    print("-------------------------------------------------------------")
+
+
+def _reset_scaling(results_dir):
+    filepath = results_dir + "/scaling.json"
+    if os.path.exists(filepath):
+        os.remove(filepath)
+
+
+def _get_scaling_factor(results_dir):
+    scaling_factor = 1.0
+    scaling_file = results_dir + "/scaling.json"
+    if os.path.exists(scaling_file):
+        with open(scaling_file, "r") as f:
+            contents = json.load(f)
+        scaling_factor = contents["scaling_factor"]
+    return scaling_factor
+
+
+def _find_benchmark(result_file, ruleset):
+    loglines, _ = parse_file(result_file, ruleset)
+    benchmark = None
+    for logline in loglines:
+        if logline.key == "submission_benchmark":
+            benchmark = logline.value["value"]
+            break
+    if benchmark is None:
+        raise ValueError("Benchmark not specified in result file")
+    return benchmark
+
+
+def _epochs_samples_to_converge(result_file, ruleset):
+    loglines, _ = parse_file(result_file, ruleset)
+    epoch_num = None
+    samples_count = None
+    for logline in loglines:
+        if logline.key == "eval_accuracy":
+            if "epoch_num" in logline.value["metadata"]:
+                epoch_num = logline.value["metadata"]["epoch_num"]
+            if "samples_count" in logline.value["metadata"]:
+                samples_count = logline.value["metadata"]["samples_count"]
+    if samples_count is not None:
+        return samples_count
+    if epoch_num is not None:
+        return epoch_num
+    raise ValueError(
+        "Not enough values specified in result file. One of ('samples_count')"
+        "or ('epoch_num') is needed"
+    )
+
+
+args = get_compute_args()
+_reset_scaling(args.benchmark_folder)
+pattern = "{folder}/result_*.txt".format(folder=args.benchmark_folder)
+result_files = glob.glob(pattern, recursive=True)
+benchmark = _find_benchmark(result_files[0], args.ruleset)
+required_runs = get_result_file_counts(args.usage)[benchmark]
+if required_runs > len(result_files):
+    print(
+        f"WARNING: Not enough runs found for an official submission."
+        f" Found: {len(result_files)}, required: {required_runs}"
+    )
+
+if args.scale:
+    rcp_checker.check_directory(
+        args.benchmark_folder,
+        args.usage,
+        args.ruleset,
+        False,
+        False,
+        rcp_file=None,
+        rcp_pass="pruned_rcps",
+        rcp_bypass=False,
+        set_scaling=True,
+    )
+
+scaling_factor = _get_scaling_factor(args.benchmark_folder)
+
+if args.is_weak_scaling:
+    scores, power_scores = result_summarizer._compute_weak_score_standalone(
+        benchmark,
+        args.system,
+        args.has_power,
+        args.benchmark_folder,
+        args.usage,
+        args.ruleset,
+    )
+    print_benchmark_info(args, benchmark)
+    print(f"Scores: {scores}")
+    if power_scores:
+        print(f"Power Scores - Energy (kJ): {power_scores}")
+else:
+    scores_track, power_scores_track, score, power_score = (
+        result_summarizer._compute_strong_score_standalone(
+            benchmark,
+            args.system,
+            args.has_power,
+            args.benchmark_folder,
+            args.usage,
+            args.ruleset,
+            return_full_scores=True,
+        )
+    )
+    print_benchmark_info(args, benchmark)
+    mean_score = 0
+    for file, s in scores_track.items():
+        epochs_samples_to_converge = _epochs_samples_to_converge(file, args.ruleset)
+        print(
+            f"Score - Time to Train (minutes) for {file}: {s}. Samples/Epochs to converge: {epochs_samples_to_converge}"
+        )
+        mean_score += s
+    mean_score /= len(result_files)
+    mean_score *= scaling_factor
+    if required_runs > len(result_files):
+        print("WARNING: Olympic scoring skipped")
+        print(f"Final score - Time to Train (minutes): {mean_score}")
+    else:
+        print(f"Final score - Time to Train (minutes): {score}")
+    if power_score:
+        mean_power = 0
+        for file, ps in power_scores_track.items():
+            print(f"Power Score - Energy (kJ) for {file}: {ps}")
+            mean_power += ps
+        mean_power /= len(result_files)
+        mean_power *= scaling_factor
+        if required_runs > len(result_files):
+            print("WARNING: Olympic scoring skipped")
+            print(f"Final score - Time to Train (minutes): {mean_power}")
+        else:
+            print(f"Power Score - Energy (kJ): {power_score}")