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

experiments/profile128k/client.py

@@ -0,0 +1,67 @@
+import argparse,json,hashlib,time,urllib.request,signal
+from pathlib import Path
+parser=argparse.ArgumentParser();parser.add_argument('--label',required=True);args=parser.parse_args()
+base='http://127.0.0.1:8000'
+key=Path('/run/secrets/qwen_api_key').read_text().strip()
+headers={'Authorization':'Bearer '+key,'Content-Type':'application/json'}
+def emit(r):print(json.dumps({'label':args.label,'time':time.time(),**r},ensure_ascii=False),flush=True)
+def request(path,payload=None):
+ req=urllib.request.Request(base+path,headers=headers,data=json.dumps(payload).encode() if payload is not None else None)
+ with urllib.request.urlopen(req,timeout=60) as res:return res.read()
+def metrics():
+ result={}
+ for l in request('/metrics').decode().splitlines():
+  if l.startswith('vllm:ple_mmap_') or l.startswith('vllm:prefix_cache_'):
+   name=l.split('{')[0].split()[0]
+   if name.endswith('_total'):result[name]=result.get(name,0)+float(l.split()[-1])
+ return result
+def deadline(*_):raise TimeoutError('Request exceeded 180 seconds')
+signal.signal(signal.SIGALRM,deadline)
+def run(name,prompt,expected=None,profile=False):
+ m=[{'role':'user','content':prompt}]
+ count=json.loads(request('/tokenize',{'model':'qwen3.8-flash-next','messages':m,'chat_template_kwargs':{'reasoning_effort':'low'}}))['count']
+ assert count+2048<=131072
+ before=metrics();payload={'model':'qwen3.8-flash-next','messages':m,'temperature':0,'seed':42,'reasoning_effort':'low','max_tokens':2048 if expected else 512,'stream':True,'stream_options':{'include_usage':True}}
+ req=urllib.request.Request(base+'/v1/chat/completions',headers=headers,data=json.dumps(payload,ensure_ascii=False).encode())
+ emit({'event':'REQUEST_START','name':name,'profile':profile,'prompt_tokens_preflight':count})
+ if profile:request('/start_profile',{})
+ t=time.monotonic();first=last=visible=None;content=reasoning='';usage={};finish=None
+ signal.alarm(180)
+ try:
+  with urllib.request.urlopen(req,timeout=180) as res:
+   for raw in res:
+    line=raw.decode().strip()
+    if not line.startswith('data: ') or line=='data: [DONE]':continue
+    x=json.loads(line[6:])
+    if x.get('error'):raise RuntimeError('Streaming error')
+    if x.get('usage'):usage=x['usage']
+    for c in x.get('choices',[]):
+     d=c.get('delta',{});a=d.get('content') or '';r=d.get('reasoning') or d.get('reasoning_content') or ''
+     if a or r:
+      now=time.monotonic();first=first if first is not None else now;last=now
+     if a:visible=visible if visible is not None else now
+     content+=a;reasoning+=r;finish=c.get('finish_reason') or finish
+ finally:
+  signal.alarm(0)
+  if profile:request('/stop_profile',{})
+ elapsed=time.monotonic()-t
+ after=metrics()
+ correct=bool(content.strip()) and (expected is None or content.strip()==expected) and finish=='stop'
+ emit({'event':'RESULT','name':name,'profile':profile,'correct':correct,'content':content,'finish_reason':finish,'usage':usage,'max_tokens':payload['max_tokens'],'prompt_sha256':hashlib.sha256(prompt.encode()).hexdigest(),'output_sha256':hashlib.sha256((reasoning+'\0'+content).encode()).hexdigest(),'ttft_s':first-t if first else None,'first_content_s':visible-t if visible else None,'elapsed_s':elapsed,'decode_tps_approx':(usage.get('completion_tokens',0)-1)/(last-first) if last and first and last>first else None,'metrics_delta':{k:after[k]-before.get(k,0) for k in after}})
+ assert correct and usage['prompt_tokens']==count,name
+ return prompt
+short='用中文写一段约150字的说明，解释数据库索引为什么能加快查询，以及它对写入有什么影响。'
+run('warmup','计算17乘19，只输出结果。','323')
+# Same prefix fixtures across configurations; phase/round identity is before the document.
+fixtures={n:json.loads(Path(f'/profiles/data/ctx-{n}-1.json').read_text())['cases'][1] for n in [8192,32768]}
+for i in range(3):
+ run(f'short-{i}',short)
+ for n,c in fixtures.items():
+  prompt=f'测试文档批次 BENCH-{i}-{n}。\n'+c['messages'][0]['content']
+  run(f'prefill-{n}-{i}',prompt,c['expected'])
+  run(f'reuse-{n}-{i}',prompt,c['expected'])
+run('trace-short',short,profile=True)
+c=fixtures[32768];prompt='测试文档批次 TRACE-32768。\n'+c['messages'][0]['content']
+run('trace-prefill-32768',prompt,c['expected'],profile=True)
+run('trace-reuse-32768',prompt,c['expected'],profile=True)
+emit({'event':'PROFILE_SUITE_PASS'})