File size: 7,466 Bytes
8d0b310
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""Concurrent API correctness/load check for ds4-server session batching.

Each case is submitted twice with the same non-zero seed. The pairs run in one
cold concurrent wave and must return identical output, even though prompt sizes
and output limits differ across cases. A fresh nonce avoids accidental reuse of
an earlier run's in-memory or disk checkpoint.
"""

import argparse
import concurrent.futures
import json
import math
import statistics
import sys
import threading
import time
import urllib.error
import urllib.request


CASES = (
    ("short-greedy", 0, 80, 0.0, 101),
    ("medium-sampled", 384, 72, 0.7, 202),
    ("long-greedy", 2304, 48, 0.0, 303),
    ("short-sampled", 32, 96, 0.65, 404),
    ("very-long-sampled", 3600, 32, 0.75, 505),
    ("medium-greedy", 1024, 64, 0.0, 606),
    ("long-sampled", 2048, 40, 0.7, 707),
    ("tiny-sampled", 0, 96, 0.6, 808),
)


def percentile(values, fraction):
    ordered = sorted(values)
    return ordered[max(0, math.ceil(len(ordered) * fraction) - 1)]


def make_payload(case, case_number, nonce, stream):
    name, filler_words, max_tokens, temperature, seed = case
    filler = (" alpha beta gamma delta" * ((filler_words + 3) // 4))
    filler = " ".join(filler.split()[:filler_words])
    prompt = (
        "This is batching test %s case %d (%s). Read the filler, then write "
        "a concise explanation of why deterministic request isolation matters. "
        "Do not quote the filler.\nFILLER:\n%s"
    ) % (nonce, case_number, name, filler)
    payload = {
        "model": "deepseek-chat",
        "messages": [{"role": "user", "content": prompt}],
        "max_tokens": max_tokens,
        "temperature": temperature,
        "top_p": 0.9,
        "seed": seed + case_number * 1000,
        "stream": stream,
    }
    if stream:
        payload["stream_options"] = {"include_usage": True}
    return name, filler_words, payload


def parse_stream(raw):
    content = []
    reasoning = []
    finish = None
    usage = None
    for line in raw.decode("utf-8", errors="replace").splitlines():
        if not line.startswith("data:"):
            continue
        data = line[5:].strip()
        if not data or data == "[DONE]":
            continue
        event = json.loads(data)
        if event.get("usage"):
            usage = event["usage"]
        choices = event.get("choices") or []
        if not choices:
            continue
        choice = choices[0]
        delta = choice.get("delta") or {}
        if delta.get("content") is not None:
            content.append(delta["content"])
        if delta.get("reasoning_content") is not None:
            reasoning.append(delta["reasoning_content"])
        if choice.get("finish_reason") is not None:
            finish = choice["finish_reason"]
    return {
        "content": "".join(content),
        "reasoning": "".join(reasoning),
        "finish": finish,
        "completion_tokens": (usage or {}).get("completion_tokens"),
    }


def post_chat(url, payload, timeout, start_event):
    start_event.wait()
    body = json.dumps(payload, separators=(",", ":")).encode("utf-8")
    req = urllib.request.Request(
        url.rstrip("/") + "/v1/chat/completions",
        data=body,
        headers={"Content-Type": "application/json"},
        method="POST",
    )
    started = time.monotonic()
    try:
        with urllib.request.urlopen(req, timeout=timeout) as response:
            raw = response.read()
    except urllib.error.HTTPError as exc:
        detail = exc.read().decode("utf-8", errors="replace")
        raise RuntimeError("HTTP %d: %s" % (exc.code, detail)) from exc
    elapsed = time.monotonic() - started

    if payload["stream"]:
        result = parse_stream(raw)
    else:
        response = json.loads(raw)
        choice = response["choices"][0]
        message = choice.get("message") or {}
        result = {
            "content": message.get("content") or "",
            "reasoning": message.get("reasoning_content") or "",
            "finish": choice.get("finish_reason"),
            "completion_tokens": (response.get("usage") or {}).get(
                "completion_tokens"
            ),
        }
    result["elapsed"] = elapsed
    return result


def comparable(result):
    return (
        result["content"],
        result["reasoning"],
        result["finish"],
        result["completion_tokens"],
    )


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--url", default="http://127.0.0.1:8000")
    parser.add_argument("--pairs", type=int, default=4)
    parser.add_argument("--workers", type=int, default=0)
    parser.add_argument("--timeout", type=float, default=1800.0)
    parser.add_argument("--stream", action="store_true")
    parser.add_argument("--nonce", default="")
    parser.add_argument(
        "--case", choices=[case[0] for case in CASES],
        help="repeat one case shape instead of cycling through mixed lengths",
    )
    args = parser.parse_args()
    if args.pairs <= 0:
        parser.error("--pairs must be positive")

    nonce = args.nonce or "cold-%d" % time.time_ns()
    requests = []
    metadata = []
    for i in range(args.pairs):
        case = next((c for c in CASES if c[0] == args.case), None)
        if case is None:
            case = CASES[i % len(CASES)]
        name, filler_words, payload = make_payload(case, i, nonce, args.stream)
        for copy in range(2):
            requests.append(payload)
            metadata.append((i, copy, name, filler_words))

    workers = args.workers or len(requests)
    workers = max(1, min(workers, len(requests)))
    start_event = threading.Event()
    wall_start = time.monotonic()
    with concurrent.futures.ThreadPoolExecutor(max_workers=workers) as executor:
        futures = [
            executor.submit(post_chat, args.url, payload, args.timeout, start_event)
            for payload in requests
        ]
        start_event.set()
        results = [future.result() for future in futures]
    wall = time.monotonic() - wall_start

    failures = 0
    for i in range(args.pairs):
        a = results[2 * i]
        b = results[2 * i + 1]
        if comparable(a) != comparable(b):
            failures += 1
            print("MISMATCH pair=%d case=%s" % (i, metadata[2 * i][2]), file=sys.stderr)
            print("  A=%s" % (json.dumps(a, ensure_ascii=True)[:1000]), file=sys.stderr)
            print("  B=%s" % (json.dumps(b, ensure_ascii=True)[:1000]), file=sys.stderr)

    latencies = [result["elapsed"] for result in results]
    known_tokens = [
        result["completion_tokens"]
        for result in results
        if result["completion_tokens"] is not None
    ]
    summary = {
        "status": "PASS" if failures == 0 else "FAIL",
        "pairs": args.pairs,
        "requests": len(requests),
        "workers": workers,
        "stream": args.stream,
        "wall_seconds": round(wall, 3),
        "latency_p50_seconds": round(statistics.median(latencies), 3),
        "latency_p95_seconds": round(percentile(latencies, 0.95), 3),
        "completion_tokens": sum(known_tokens) if known_tokens else None,
        "completion_tokens_per_second": (
            round(sum(known_tokens) / wall, 2) if known_tokens else None
        ),
        "nonce": nonce,
    }
    print(json.dumps(summary, sort_keys=True))
    return 0 if failures == 0 else 1


if __name__ == "__main__":
    raise SystemExit(main())