|
| 1 | +import random |
| 2 | +from typing import Dict, List, Tuple |
| 3 | + |
| 4 | +from vllm_router.routers.routing_logic import RoundRobinRouter |
| 5 | + |
| 6 | + |
| 7 | +class EndpointInfo: |
| 8 | + def __init__(self, url: str): |
| 9 | + self.url = url |
| 10 | + |
| 11 | + |
| 12 | +class RequestStats: |
| 13 | + def __init__(self, qps: float): |
| 14 | + self.qps = qps |
| 15 | + |
| 16 | + |
| 17 | +class Request: |
| 18 | + def __init__(self, headers: Dict[str, str]): |
| 19 | + self.headers = headers |
| 20 | + |
| 21 | + |
| 22 | +class EngineStats: |
| 23 | + def __init__(self): |
| 24 | + return |
| 25 | + |
| 26 | + |
| 27 | +def generate_request_args( |
| 28 | + num_endpoints: int, qps_range: int = 0 |
| 29 | +) -> Tuple[List[EndpointInfo], Dict[str, EngineStats], Dict[str, RequestStats]]: |
| 30 | + endpoints = [ |
| 31 | + EndpointInfo( |
| 32 | + url=f"{endpoint_index}", |
| 33 | + ) |
| 34 | + for endpoint_index in range(num_endpoints) |
| 35 | + ] |
| 36 | + engine_stats = { |
| 37 | + f"{endpoint_index}": EngineStats() for endpoint_index in range(num_endpoints) |
| 38 | + } |
| 39 | + request_stats = { |
| 40 | + f"{endpoint_index}": RequestStats(qps=random.uniform(0, qps_range)) |
| 41 | + for endpoint_index in range(num_endpoints) |
| 42 | + } |
| 43 | + return endpoints, engine_stats, request_stats |
| 44 | + |
| 45 | + |
| 46 | +def generate_request(request_type="http") -> Request: |
| 47 | + return Request({"type": request_type}) |
| 48 | + |
| 49 | + |
| 50 | +def test_roundrobin_logic( |
| 51 | + dynamic_discoveries: int = 10, max_endpoints: int = 1000, max_requests: int = 10000 |
| 52 | +): |
| 53 | + """ |
| 54 | + Ensure that all active urls have roughly same number of requests (difference at most 1) |
| 55 | + """ |
| 56 | + router = RoundRobinRouter() |
| 57 | + |
| 58 | + def _fixed_router_check(num_endpoints: int, num_requests: int) -> bool: |
| 59 | + # Make num_requests requests to the router and check even output distribution |
| 60 | + endpoints, engine_stats, request_stats = generate_request_args(num_endpoints) |
| 61 | + output_distribution = {} |
| 62 | + for request_idx in range(num_requests): |
| 63 | + request = generate_request() |
| 64 | + url = router.route_request(endpoints, engine_stats, request_stats, request) |
| 65 | + output_distribution[url] = output_distribution.get(url, 0) + 1 |
| 66 | + request_counts = output_distribution.values() |
| 67 | + return max(request_counts) - min(request_counts) <= 1 |
| 68 | + |
| 69 | + for _ in range(dynamic_discoveries): |
| 70 | + num_endpoints = random.randint(1, max_endpoints) |
| 71 | + num_requests = random.randint(1, max_requests) |
| 72 | + # Perform router check |
| 73 | + res = _fixed_router_check(num_endpoints, num_requests) |
| 74 | + assert res |
0 commit comments