基于AI Agent开发范式的国产GPU大模型推理算子库优化/baselines/flashinfer_baseline/summarize_results.py

@@ -0,0 +1,135 @@
+"""Summarize FlashInfer benchmark CSV files into a Markdown report."""
+
+import argparse
+import csv
+from collections import Counter, defaultdict
+from pathlib import Path
+
+
+NUMERIC_COLUMNS = ("time_ms", "bandwidth_GB_s", "tflops")
+
+
+def parse_float(value):
+    try:
+        return float(value)
+    except (TypeError, ValueError):
+        return None
+
+
+def load_records(paths):
+    records = []
+    for path in paths:
+        with path.open(newline="") as f:
+            reader = csv.DictReader(f)
+            for row in reader:
+                row["_source"] = str(path)
+                records.append(row)
+    return records
+
+
+def group_by_api(records):
+    grouped = defaultdict(list)
+    for record in records:
+        grouped[record.get("api", "unknown")].append(record)
+    return dict(grouped)
+
+
+def best_record(records, column, reverse=True):
+    candidates = []
+    for record in records:
+        value = parse_float(record.get(column))
+        if value is not None:
+            candidates.append((value, record))
+    if not candidates:
+        return None
+    return sorted(candidates, key=lambda item: item[0], reverse=reverse)[0]
+
+
+def describe_case(record):
+    case_keys = [
+        "batch_size",
+        "seq_len",
+        "seq_len_q",
+        "seq_len_kv",
+        "num_heads",
+        "num_qo_heads",
+        "num_kv_heads",
+        "head_dim",
+        "head_dim_qk",
+        "head_dim_vo",
+        "head_dim_ckv",
+        "head_dim_kpe",
+    ]
+    parts = [f"{key}={record[key]}" for key in case_keys if record.get(key)]
+    return ", ".join(parts) if parts else "n/a"
+
+
+def build_markdown(records):
+    lines = [
+        "# FlashInfer Benchmark Summary",
+        "",
+        f"- CSV files: {len({record['_source'] for record in records})}",
+        f"- Total rows: {len(records)}",
+        "",
+    ]
+
+    for api, api_records in sorted(group_by_api(records).items()):
+        statuses = Counter(record.get("status", "ok") or "ok" for record in api_records)
+        lines.extend(
+            [
+                f"## {api}",
+                "",
+                f"- Rows: {len(api_records)}",
+                f"- Status: {', '.join(f'{key}={value}' for key, value in sorted(statuses.items()))}",
+            ]
+        )
+
+        for column in NUMERIC_COLUMNS:
+            best = best_record(api_records, column, reverse=(column != "time_ms"))
+            if best is None:
+                continue
+            value, record = best
+            lines.append(f"- Best {column}: {value:.6g} ({describe_case(record)})")
+
+        failures = [
+            record
+            for record in api_records
+            if record.get("status") and record.get("status") != "ok"
+        ]
+        for failure in failures[:5]:
+            lines.append(
+                f"- Failure: {describe_case(failure)}: "
+                f"{failure.get('error_type', 'error')} {failure.get('error', '')}".strip()
+            )
+        if len(failures) > 5:
+            lines.append(f"- Additional failures omitted: {len(failures) - 5}")
+        lines.append("")
+
+    return "\n".join(lines)
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("csv", nargs="+", type=Path, help="Benchmark CSV files.")
+    parser.add_argument(
+        "--output",
+        type=Path,
+        default=Path("flashinfer_benchmark_summary.md"),
+        help="Markdown summary output path.",
+    )
+    return parser.parse_args()
+
+
+def main():
+    args = parse_args()
+    records = load_records(args.csv)
+    if not records:
+        raise ValueError("no benchmark rows found")
+    markdown = build_markdown(records)
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    args.output.write_text(markdown, encoding="utf-8")
+    print(f"Summary saved to {args.output}")
+
+
+if __name__ == "__main__":
+    main()