File size: 12,832 Bytes
510314e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 | """Contatore delle chiamate al motore (banco ALB v2, BRIEF §5): un passante HTTP fra chi chiama (banco per A/T/B,
mnemonic-proxy per D/E) e il motore (Strata o motore finto). Ogni POST /v1/chat/completions che arriva al motore
passa di qui UNA volta e viene registrato una volta: giri interni del proxy, note di passaggio, retry K, rigioco,
compattazioni di B, risposta finale. Nessuna modifica al corpo o alla risposta (passaggio byte per byte; lo stream
SSE viene inoltrato riga per riga mentre lo si legge).
Per chiamata: kind, max_tokens, reasoning_budget_tokens, enable_thinking, strumenti offerti, messaggi, prompt /
generazione / cache (usage), timings del motore, TTFT e durata visti dal contatore, finish_reason, caratteri di
content e ragionamento, chiamate a strumenti, stato HTTP, e il controllo a posteriori
prompt_tokens + max_tokens + margine <= W (margin_violation).
kind: header X-Bench-Kind se presente (chiamate dirette del banco); altrimenti (chiamate del proxy):
replay max_tokens == 1 (rigioco della storia)
notes un messaggio contiene il marcatore delle note di passaggio
round l'ultimo messaggio è un risultato di strumento (giro interno: dopo recall/answerfix)
final il resto (richiesta finale inoltrata, eventuale retry K senza strumenti)
"""
from __future__ import annotations
import http.client
import json
import threading
import time
import urllib.parse
from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
NOTES_MARKS = ("[strata-context: handoff notes request]", "[strata-context: richiesta note di passaggio]")
HOP = {"connection", "keep-alive", "transfer-encoding", "te", "trailer", "upgrade", "proxy-connection", "content-length"}
def _kind(body: dict, hdr: str | None) -> str:
if hdr:
return hdr
if int(body.get("max_tokens") or body.get("max_completion_tokens") or 0) == 1:
return "replay"
msgs = body.get("messages") or []
for m in msgs:
c = m.get("content")
c = c if isinstance(c, str) else json.dumps(c, ensure_ascii=False) if c else ""
if any(k in c for k in NOTES_MARKS):
return "notes"
if msgs and msgs[-1].get("role") == "tool":
return "round"
return "final"
class Meter:
def __init__(self, upstream: str, log_path, window: int, margin: int = 64, port: int = 0):
u = urllib.parse.urlparse(upstream)
self.host, self.uport = u.hostname, u.port or 80
self.window, self.margin = int(window), int(margin)
self.logf = open(log_path, "a")
self.lock = threading.Lock()
self.n = 0
meter = self
class H(BaseHTTPRequestHandler):
protocol_version = "HTTP/1.1"
def log_message(self, *a):
pass
def _fwd_simple(self, method, data=None):
conn = http.client.HTTPConnection(meter.host, meter.uport, timeout=3600)
hdrs = {k: v for k, v in self.headers.items() if k.lower() not in HOP and k.lower() != "host"}
conn.request(method, self.path, body=data, headers=hdrs)
r = conn.getresponse()
return conn, r
def do_GET(self):
conn, r = self._fwd_simple("GET")
d = r.read()
self.send_response(r.status)
for k, v in r.getheaders():
if k.lower() not in HOP:
self.send_header(k, v)
self.send_header("Content-Length", str(len(d)))
self.end_headers()
self.wfile.write(d)
conn.close()
def do_POST(self):
n = int(self.headers.get("Content-Length") or 0)
data = self.rfile.read(n)
if not self.path.startswith("/v1/chat/completions"):
conn, r = self._fwd_simple("POST", data)
d = r.read()
self.send_response(r.status)
for k, v in r.getheaders():
if k.lower() not in HOP:
self.send_header(k, v)
self.send_header("Content-Length", str(len(d)))
self.end_headers()
self.wfile.write(d)
conn.close()
return
try:
body = json.loads(data or b"{}")
except ValueError:
body = {}
ctk = body.get("chat_template_kwargs") or {}
rec = {"t_start": time.time(), "kind": _kind(body, self.headers.get("X-Bench-Kind")),
"max_tokens": int(body.get("max_tokens") or body.get("max_completion_tokens") or 0),
"reasoning_budget_tokens": body.get("reasoning_budget_tokens"),
"enable_thinking": ctk.get("enable_thinking"), "stream": bool(body.get("stream")),
"n_messages": len(body.get("messages") or []),
"tools": [(t.get("function") or {}).get("name") for t in body.get("tools") or []]}
t0 = time.time()
conn, r = self._fwd_simple("POST", data)
rec["http_status"] = r.status
ttft, usage, timings, finish = None, None, None, None
content, reasoning, calls = 0, 0, {}
ctype = r.getheader("Content-Type") or ""
if r.status >= 400 or "event-stream" not in ctype:
d = r.read()
self.send_response(r.status)
for k, v in r.getheaders():
if k.lower() not in HOP:
self.send_header(k, v)
self.send_header("Content-Length", str(len(d)))
self.end_headers()
self.wfile.write(d)
ttft = time.time() - t0
try:
j = json.loads(d)
if r.status >= 400:
rec["error"] = json.dumps(j.get("error") or j)[:500]
else:
ch = (j.get("choices") or [{}])[0]
msg = ch.get("message") or {}
usage, timings, finish = j.get("usage"), j.get("timings"), ch.get("finish_reason")
content = len(msg.get("content") or "")
reasoning = len(msg.get("reasoning_content") or "")
for i, c in enumerate(msg.get("tool_calls") or []):
calls[i] = (c.get("function") or {}).get("name")
except ValueError:
rec["error"] = d[:500].decode("utf-8", "replace")
else:
self.send_response(r.status)
for k, v in r.getheaders():
if k.lower() not in HOP:
self.send_header(k, v)
self.send_header("Connection", "close")
self.end_headers()
while True:
line = r.readline()
if not line:
break
self.wfile.write(line)
s = line.strip()
if s.startswith(b"data:"):
p = s[5:].strip()
if p == b"[DONE]":
self.wfile.flush()
continue
try:
c = json.loads(p)
except ValueError:
continue
usage = c.get("usage") or usage
timings = c.get("timings") or timings
for ch in c.get("choices") or []:
dl = ch.get("delta") or {}
if ttft is None and (dl.get("content") or dl.get("reasoning_content")
or dl.get("tool_calls")):
ttft = time.time() - t0
content += len(dl.get("content") or "")
reasoning += len(dl.get("reasoning_content") or "")
for tc in dl.get("tool_calls") or []:
nm = (tc.get("function") or {}).get("name")
k = tc.get("index", len(calls))
calls[k] = (calls.get(k) or "") + (nm or "")
finish = ch.get("finish_reason") or finish
elif not s:
self.wfile.flush()
self.wfile.flush()
self.close_connection = True
conn.close()
u = usage or {}
pt = u.get("prompt_tokens")
rec.update(wall_s=time.time() - t0, ttft_s=ttft, finish_reason=finish, usage=usage, timings=timings,
prompt_tokens=pt, completion_tokens=u.get("completion_tokens"),
cached_tokens=((u.get("prompt_tokens_details") or {}).get("cached_tokens")
if isinstance(u.get("prompt_tokens_details"), dict) else None),
content_chars=content, reasoning_chars=reasoning,
tool_calls=[calls[k] for k in sorted(calls)])
if pt is not None:
rec["window_slack"] = meter.window - (int(pt) + rec["max_tokens"] + meter.margin)
rec["margin_violation"] = rec["window_slack"] < 0
with meter.lock:
meter.n += 1
rec["seq"] = meter.n
meter.logf.write(json.dumps(rec, ensure_ascii=False) + "\n")
meter.logf.flush()
self.srv = ThreadingHTTPServer(("127.0.0.1", port), H)
self.srv.daemon_threads = True
self.port = self.srv.server_address[1]
self.url = "http://127.0.0.1:%d" % self.port
self.th = threading.Thread(target=self.srv.serve_forever, daemon=True)
self.th.start()
def close(self):
self.srv.shutdown()
self.srv.server_close()
self.logf.close()
def summarize(path) -> dict:
"""Costi completi da engine_calls.jsonl: totali e per kind (ogni chiamata al motore contata una volta)."""
tot = {"n_calls": 0, "prompt_tokens": 0, "completion_tokens": 0, "cached_tokens": 0, "prompt_n": 0, "cache_n": 0,
"prompt_ms": 0.0, "predicted_ms": 0.0, "wall_s": 0.0, "max_prompt_tokens": 0, "finish_length": 0,
"empty_output": 0, "http_errors": 0, "margin_violations": 0, "min_window_slack": None,
"max_tokens_sum": 0, "reasoning_budget_given": 0}
by = {}
try:
lines = open(path).read().splitlines()
except OSError:
lines = []
for l in lines:
c = json.loads(l)
for t in (tot, by.setdefault(c["kind"], {k: (0 if not isinstance(v, float) else 0.0) if v is not None else None
for k, v in tot.items()})):
t["n_calls"] += 1
t["prompt_tokens"] += int(c.get("prompt_tokens") or 0)
t["completion_tokens"] += int(c.get("completion_tokens") or 0)
t["cached_tokens"] += int(c.get("cached_tokens") or 0)
tm = c.get("timings") or {}
t["prompt_n"] += int(tm.get("prompt_n") or 0)
t["cache_n"] += int(tm.get("cache_n") or 0)
t["prompt_ms"] += float(tm.get("prompt_ms") or 0.0)
t["predicted_ms"] += float(tm.get("predicted_ms") or 0.0)
t["wall_s"] += float(c.get("wall_s") or 0.0)
t["max_prompt_tokens"] = max(t["max_prompt_tokens"], int(c.get("prompt_tokens") or 0))
t["finish_length"] += c.get("finish_reason") == "length"
t["empty_output"] += (c.get("http_status") == 200 and not c.get("content_chars")
and not c.get("tool_calls") and c.get("max_tokens", 0) > 1)
t["http_errors"] += int(c.get("http_status") or 0) >= 400
t["margin_violations"] += bool(c.get("margin_violation"))
if c.get("window_slack") is not None:
t["min_window_slack"] = c["window_slack"] if t["min_window_slack"] is None else \
min(t["min_window_slack"], c["window_slack"])
t["max_tokens_sum"] += int(c.get("max_tokens") or 0)
t["reasoning_budget_given"] += c.get("reasoning_budget_tokens") is not None
tot["by_kind"] = by
return tot
|