|
| 1 | +#!/usr/bin/env python3 |
| 2 | +"""Run one cold-prefix request and prove that a remote prefill worker served it.""" |
| 3 | +from __future__ import annotations |
| 4 | + |
| 5 | +import argparse |
| 6 | +import json |
| 7 | +import secrets |
| 8 | +import time |
| 9 | +import urllib.request |
| 10 | + |
| 11 | + |
| 12 | +def _get_json(url: str): |
| 13 | + with urllib.request.urlopen(url, timeout=5) as response: |
| 14 | + return json.load(response) |
| 15 | + |
| 16 | + |
| 17 | +def _delta(before: dict, after: dict, key: str) -> int: |
| 18 | + return int(after.get(key, 0)) - int(before.get(key, 0)) |
| 19 | + |
| 20 | + |
| 21 | +def main() -> int: |
| 22 | + parser = argparse.ArgumentParser() |
| 23 | + parser.add_argument("--address", default="127.0.0.1:51051") |
| 24 | + parser.add_argument("--dashboard", default="http://127.0.0.1:8090") |
| 25 | + parser.add_argument("--tokenizer-id", required=True) |
| 26 | + parser.add_argument("--minimum-prefix-tokens", type=int, default=128) |
| 27 | + args = parser.parse_args() |
| 28 | + |
| 29 | + from kakeya import Client |
| 30 | + from transformers import AutoTokenizer |
| 31 | + |
| 32 | + from scripts.chat_grpc import _resolve_eos_token_ids |
| 33 | + |
| 34 | + tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_id) |
| 35 | + nonce = secrets.token_hex(8) |
| 36 | + sentence = ( |
| 37 | + "Kakeya remote prefill verification context. " |
| 38 | + f"Unique run {nonce}. " |
| 39 | + ) |
| 40 | + text = sentence |
| 41 | + token_ids = tokenizer.encode(text, add_special_tokens=True) |
| 42 | + while len(token_ids) < args.minimum_prefix_tokens: |
| 43 | + text += sentence |
| 44 | + token_ids = tokenizer.encode(text, add_special_tokens=True) |
| 45 | + |
| 46 | + nodes = _get_json(f"{args.dashboard}/v1/network/nodes") |
| 47 | + workers = [ |
| 48 | + node for node in nodes |
| 49 | + if node.get("prefill_worker") and node.get("status") == "online" |
| 50 | + ] |
| 51 | + if not workers: |
| 52 | + print(json.dumps({ |
| 53 | + "ok": False, |
| 54 | + "reason": "no online prefill worker capability", |
| 55 | + "nodes": nodes, |
| 56 | + }, indent=2)) |
| 57 | + return 2 |
| 58 | + |
| 59 | + before = _get_json(f"{args.dashboard}/v1/network/prefill") |
| 60 | + started = time.perf_counter() |
| 61 | + with Client(args.address) as client: |
| 62 | + with client.create_session( |
| 63 | + eos_token_ids=_resolve_eos_token_ids(tokenizer), |
| 64 | + client_label="remote-prefill-e2e", |
| 65 | + ) as session: |
| 66 | + session.append(token_ids) |
| 67 | + list(session.generate(max_tokens=1)) |
| 68 | + elapsed = time.perf_counter() - started |
| 69 | + after = _get_json(f"{args.dashboard}/v1/network/prefill") |
| 70 | + |
| 71 | + result = { |
| 72 | + "ok": ( |
| 73 | + _delta(before, after, "remote_jobs") >= 1 |
| 74 | + and _delta(before, after, "remote_hits") >= 1 |
| 75 | + and _delta(before, after, "tokens_reused") |
| 76 | + >= args.minimum_prefix_tokens |
| 77 | + ), |
| 78 | + "worker_nodes": [node["id"] for node in workers], |
| 79 | + "prefix_tokens": len(token_ids), |
| 80 | + "wall_seconds": elapsed, |
| 81 | + "delta": { |
| 82 | + key: _delta(before, after, key) |
| 83 | + for key in ( |
| 84 | + "remote_jobs", |
| 85 | + "remote_hits", |
| 86 | + "tokens_reused", |
| 87 | + "tokens_computed", |
| 88 | + "bytes_received", |
| 89 | + "remote_job_failures", |
| 90 | + "fallbacks", |
| 91 | + ) |
| 92 | + }, |
| 93 | + "before": before, |
| 94 | + "after": after, |
| 95 | + } |
| 96 | + print(json.dumps(result, indent=2)) |
| 97 | + return 0 if result["ok"] else 1 |
| 98 | + |
| 99 | + |
| 100 | +if __name__ == "__main__": |
| 101 | + raise SystemExit(main()) |
0 commit comments