|
| 1 | +from fastapi import FastAPI, WebSocket, WebSocketDisconnect, Request, Query |
| 2 | +from threading import Lock |
| 3 | +from typing import Dict |
| 4 | +from fastapi.responses import JSONResponse |
| 5 | +from lightllm.utils.log_utils import init_logger |
| 6 | +from ..pd_io_struct import PD_Master_Obj |
| 7 | +import base64 |
| 8 | +import pickle |
| 9 | +import os |
| 10 | +import requests |
| 11 | + |
| 12 | +logger = init_logger(__name__) |
| 13 | +app = FastAPI() |
| 14 | + |
| 15 | +registered_pd_master_objs: Dict[str, PD_Master_Obj] = {} |
| 16 | +registered_pd_master_obj_lock = Lock() |
| 17 | + |
| 18 | +global_req_id = 0 |
| 19 | +global_req_id_lock = Lock() |
| 20 | + |
| 21 | + |
| 22 | +@app.get("/liveness") |
| 23 | +@app.post("/liveness") |
| 24 | +def liveness(): |
| 25 | + return {"status": "ok"} |
| 26 | + |
| 27 | + |
| 28 | +@app.get("/readiness") |
| 29 | +@app.post("/readiness") |
| 30 | +def readiness(): |
| 31 | + return {"status": "ok"} |
| 32 | + |
| 33 | + |
| 34 | +@app.get("/healthz", summary="Check server health") |
| 35 | +@app.get("/health", summary="Check server health") |
| 36 | +@app.head("/health", summary="Check server health") |
| 37 | +async def healthcheck(request: Request): |
| 38 | + return JSONResponse({"message": "Ok"}, status_code=200) |
| 39 | + |
| 40 | + |
| 41 | +@app.websocket("/pd_master_register") |
| 42 | +async def websocket_endpoint(websocket: WebSocket): |
| 43 | + await websocket.accept() |
| 44 | + client_ip, client_port = websocket.client |
| 45 | + logger.info(f"ws connected from IP: {client_ip}, Port: {client_port}") |
| 46 | + registered_pd_master_obj: PD_Master_Obj = pickle.loads(await websocket.receive_bytes()) |
| 47 | + logger.info(f"recieved registered_pd_master_obj {registered_pd_master_obj}") |
| 48 | + with registered_pd_master_obj_lock: |
| 49 | + registered_pd_master_objs[registered_pd_master_obj.node_id] = registered_pd_master_obj |
| 50 | + |
| 51 | + try: |
| 52 | + while True: |
| 53 | + data = await websocket.receive_text() |
| 54 | + assert data == "heartbeat" |
| 55 | + except (WebSocketDisconnect, Exception, RuntimeError) as e: |
| 56 | + logger.error(f"registered_pd_master_obj {registered_pd_master_obj} has error {str(e)}") |
| 57 | + logger.exception(str(e)) |
| 58 | + finally: |
| 59 | + logger.error(f"registered_pd_master_obj {registered_pd_master_obj} removed") |
| 60 | + with registered_pd_master_obj_lock: |
| 61 | + registered_pd_master_objs.pop(registered_pd_master_obj.node_id, None) |
| 62 | + return |
| 63 | + |
| 64 | + |
| 65 | +@app.get("/registered_objects") |
| 66 | +async def get_registered_objects(): |
| 67 | + with registered_pd_master_obj_lock: |
| 68 | + serialized_data = pickle.dumps(registered_pd_master_objs) |
| 69 | + base64_encoded = base64.b64encode(serialized_data).decode("utf-8") |
| 70 | + return {"data": base64_encoded} |
| 71 | + |
| 72 | + |
| 73 | +@app.get("/allocate_global_unique_id_range") |
| 74 | +async def allocate_global_id_range(): |
| 75 | + """ |
| 76 | + Allocate a global ID range for the requesting client without requiring parameters. |
| 77 | +
|
| 78 | + Returns: |
| 79 | + dict: A dictionary containing the start and end of the allocated ID range. |
| 80 | +
|
| 81 | + Example HTTP client usage: |
| 82 | + ```python |
| 83 | + response = requests.get("http://<server_address>/allocate_global_unique_id_range") |
| 84 | + print(response.json()) |
| 85 | + ``` |
| 86 | + """ |
| 87 | + global global_req_id |
| 88 | + range_size = 800000 |
| 89 | + with global_req_id_lock: |
| 90 | + if global_req_id + range_size > 2 ** 63 - 1: |
| 91 | + global_req_id = 0 |
| 92 | + start_id = global_req_id |
| 93 | + global_req_id += range_size |
| 94 | + end_id = global_req_id |
| 95 | + |
| 96 | + return {"start_id": start_id, "end_id": end_id} |
0 commit comments