docs(perf): profile CUDA graph coverage and PLE costs on Spark

experiments/profile128k/summarize_bench.py

@@ -0,0 +1,20 @@
+import json,statistics,sys
+from pathlib import Path
+result={}
+for fn in sys.argv[1:]:
+ rows=[]
+ for line in Path(fn).read_text().splitlines():
+  try:r=json.loads(line)
+  except ValueError:continue
+  if r.get('event')=='RESULT':rows.append(r)
+ label=rows[0]['label'];groups={}
+ for name in ['short','prefill-8192','reuse-8192','prefill-32768','reuse-32768']:
+  rr=[r for r in rows if not r['profile'] and r['name'].rsplit('-',1)[0]==name]
+  if not rr:continue
+  groups[name]={'n':len(rr),'correct':sum(r['correct'] for r in rr)}
+  for k in ['ttft_s','elapsed_s','decode_tps_approx']:
+   values=[r[k] for r in rr];groups[name][k]={'median':statistics.median(values),'values':values}
+  groups[name]['output_tokens']=[r['usage']['completion_tokens'] for r in rr]
+  groups[name]['reasoning_tokens']=[r['usage'].get('completion_tokens_details',{}).get('reasoning_tokens') for r in rr]
+ result[label]={'total_requests':len(rows),'correct':sum(r['correct'] for r in rows),'groups':groups}
+print(json.dumps(result,indent=2))