|
| 1 | +import asyncio |
| 2 | +import heapq |
| 3 | +from abc import ABC, abstractmethod |
| 4 | +from typing import List, Union |
| 5 | + |
| 6 | +from tensorrt_llm.serve.openai_protocol import (ChatCompletionRequest, |
| 7 | + CompletionRequest) |
| 8 | + |
| 9 | + |
| 10 | +def get_request_num_tokens( |
| 11 | + request: Union[CompletionRequest, ChatCompletionRequest]) -> int: |
| 12 | + if request.disaggregated_params.request_type == "context_only": |
| 13 | + if isinstance(request, ChatCompletionRequest): |
| 14 | + raise ValueError( |
| 15 | + "LoadBalancing router with tokens doesn't support ChatCompletionRequest yet" |
| 16 | + ) |
| 17 | + |
| 18 | + if isinstance(request.prompt, str) or \ |
| 19 | + (isinstance(request.prompt, list) and isinstance(request.prompt[0], int)): |
| 20 | + prompts = [request.prompt] |
| 21 | + else: |
| 22 | + prompts = request.prompt |
| 23 | + |
| 24 | + num_tokens = sum(len(prompt) for prompt in prompts) |
| 25 | + elif request.disaggregated_params.request_type == "generation_only": |
| 26 | + raise ValueError( |
| 27 | + "LoadBalancing router with tokens doesn't support generation_only requests" |
| 28 | + ) |
| 29 | + else: |
| 30 | + raise ValueError( |
| 31 | + f"Unsupported request type: {request.disaggregated_params.request_type}" |
| 32 | + ) |
| 33 | + |
| 34 | + return num_tokens |
| 35 | + |
| 36 | + |
| 37 | +class ServerState: |
| 38 | + |
| 39 | + def __init__(self, server: str, use_tokens: bool = False): |
| 40 | + self._server = server |
| 41 | + self._num_active_requests = 0 |
| 42 | + self._num_active_tokens = 0 |
| 43 | + self._use_tokens = use_tokens |
| 44 | + self._lock = asyncio.Lock() |
| 45 | + |
| 46 | + async def increment_load(self, request: Union[CompletionRequest, |
| 47 | + ChatCompletionRequest]): |
| 48 | + num_tokens = get_request_num_tokens(request) if self._use_tokens else 0 |
| 49 | + async with self._lock: |
| 50 | + self._num_active_requests += 1 |
| 51 | + self._num_active_tokens += num_tokens |
| 52 | + |
| 53 | + async def decrement_load(self, request: Union[CompletionRequest, |
| 54 | + ChatCompletionRequest]): |
| 55 | + num_tokens = get_request_num_tokens(request) if self._use_tokens else 0 |
| 56 | + async with self._lock: |
| 57 | + self._num_active_requests -= 1 |
| 58 | + self._num_active_tokens -= num_tokens |
| 59 | + |
| 60 | + |
| 61 | +class Router(ABC): |
| 62 | + |
| 63 | + def __init__(self, servers: List[str] = None): |
| 64 | + self._servers = servers |
| 65 | + |
| 66 | + @abstractmethod |
| 67 | + async def get_next_server( |
| 68 | + self, request: Union[CompletionRequest, |
| 69 | + ChatCompletionRequest]) -> str: |
| 70 | + pass |
| 71 | + |
| 72 | + @abstractmethod |
| 73 | + async def finish_request(self, request: Union[CompletionRequest, |
| 74 | + ChatCompletionRequest]): |
| 75 | + pass |
| 76 | + |
| 77 | + |
| 78 | +class RoundRobinRouter(Router): |
| 79 | + |
| 80 | + def __init__(self, servers: List[str] = None): |
| 81 | + super().__init__(servers) |
| 82 | + self._server_idx = 0 |
| 83 | + |
| 84 | + async def get_next_server( |
| 85 | + self, request: Union[CompletionRequest, |
| 86 | + ChatCompletionRequest]) -> str: |
| 87 | + server = self._servers[self._server_idx] |
| 88 | + self._server_idx = (self._server_idx + 1) % len(self._servers) |
| 89 | + return server |
| 90 | + |
| 91 | + async def finish_request(self, request: Union[CompletionRequest, |
| 92 | + ChatCompletionRequest]): |
| 93 | + pass |
| 94 | + |
| 95 | + |
| 96 | +class LoadBalancingRouter(Router): |
| 97 | + |
| 98 | + def __init__(self, servers: List[str] = None, use_tokens: bool = False): |
| 99 | + super().__init__(servers) |
| 100 | + self._lock = asyncio.Lock() |
| 101 | + # Load map between servers and their number of tokens processed |
| 102 | + self._server_state = {} |
| 103 | + self._server_load_heap = [] |
| 104 | + |
| 105 | + # Routing table to map requests to servers |
| 106 | + self._req_routing_table = {} |
| 107 | + |
| 108 | + self._use_tokens = use_tokens |
| 109 | + self._init_heap() |
| 110 | + |
| 111 | + def _init_heap(self): |
| 112 | + for server in self._servers: |
| 113 | + self._server_state[server] = ServerState(server, self._use_tokens) |
| 114 | + heapq.heappush(self._server_load_heap, |
| 115 | + (self._get_server_load(server), server)) |
| 116 | + |
| 117 | + async def get_next_server( |
| 118 | + self, request: Union[CompletionRequest, |
| 119 | + ChatCompletionRequest]) -> str: |
| 120 | + async with self._lock: |
| 121 | + server = heapq.heappop(self._server_load_heap)[1] |
| 122 | + await self._server_state[server].increment_load(request) |
| 123 | + heapq.heappush(self._server_load_heap, |
| 124 | + (self._get_server_load(server), server)) |
| 125 | + |
| 126 | + self._req_routing_table[id(request)] = server |
| 127 | + |
| 128 | + return server |
| 129 | + |
| 130 | + def _get_server_load(self, server): |
| 131 | + return self._server_state[server]._num_active_tokens if self._use_tokens \ |
| 132 | + else self._server_state[server]._num_active_requests |
| 133 | + |
| 134 | + async def finish_request(self, request: Union[CompletionRequest, |
| 135 | + ChatCompletionRequest]): |
| 136 | + async with self._lock: |
| 137 | + server = self._req_routing_table[id(request)] |
| 138 | + await self._server_state[server].decrement_load(request) |
| 139 | + heapq.heappush(self._server_load_heap, |
| 140 | + (self._get_server_load(server), server)) |
| 141 | + del self._req_routing_table[id(request)] |
| 142 | + |
| 143 | + |
| 144 | +def create_router(router_type: str, servers: List[str]) -> Router: |
| 145 | + """ |
| 146 | + Factory function to create different types of router instances. |
| 147 | +
|
| 148 | + Args: |
| 149 | + router_type (str): Type of router to create. Supported values: |
| 150 | + - "round_robin": Creates a RoundRobinRouter |
| 151 | + - "requests_load_balancing": Creates a LoadBalancingRouter, which balances requests across instances |
| 152 | + - "tokens_load_balancing": Creates a LoadBalancingRouter, which balances tokens across instances |
| 153 | + servers: List of server URLs |
| 154 | +
|
| 155 | + Returns: |
| 156 | + Router: An instance of the requested router type |
| 157 | +
|
| 158 | + Raises: |
| 159 | + ValueError: If an unsupported router type is provided |
| 160 | + """ |
| 161 | + |
| 162 | + router_map = { |
| 163 | + "round_robin": RoundRobinRouter, |
| 164 | + "requests_load_balancing": LoadBalancingRouter, |
| 165 | + "tokens_load_balancing": LoadBalancingRouter |
| 166 | + } |
| 167 | + |
| 168 | + router_class = router_map.get(router_type.lower()) |
| 169 | + if router_class is None: |
| 170 | + raise ValueError(f"Unsupported router type: {router_type}. " |
| 171 | + f"Supported types are: {list(router_map.keys())}") |
| 172 | + |
| 173 | + if router_type.endswith("load_balancing"): |
| 174 | + use_tokens = True if router_type.startswith("tokens") else False |
| 175 | + return router_class(servers, use_tokens=use_tokens) |
| 176 | + else: |
| 177 | + return router_class(servers) |
0 commit comments