[PD] Route PD server warmup to every DP rank (#30748)
Co-authored-by: weireweire <20922698+weireweire@users.noreply.github.com>
This commit is contained in:
@@ -21,6 +21,7 @@ import asyncio
|
||||
import dataclasses
|
||||
import logging
|
||||
import os
|
||||
import ssl
|
||||
import tempfile
|
||||
import threading
|
||||
import time
|
||||
@@ -39,6 +40,7 @@ from typing import (
|
||||
Union,
|
||||
)
|
||||
|
||||
import aiohttp
|
||||
import numpy as np
|
||||
import requests
|
||||
import uvicorn
|
||||
@@ -2013,6 +2015,46 @@ def _admin_api_key_missing_response(
|
||||
MINIMUM_PNG_PICTURE_BASE64 = "iVBORw0KGgoAAAANSUhEUgAAACAAAAAgCAYAAABzenr0AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAbUlEQVRYhe3VsQ2AMAxE0Y/lIgNQULD/OqyCMgCihCKSG4yRuKuiNH6JLsoEbMACOGBcua9HOR7Y6w6swBwMy0qLTpkeI77qdEBpBFAHBBDAGH8WrwJKI4AAegUCfAKgEgpQDvh3CR3oQCuav58qlAw73kKCSgAAAABJRU5ErkJggg=="
|
||||
|
||||
|
||||
async def _send_disaggregation_warmup_requests(
|
||||
server_args: ServerArgs,
|
||||
url: str,
|
||||
headers: Dict[str, str],
|
||||
ssl_verify: Union[bool, str],
|
||||
timeout: int,
|
||||
) -> List[int]:
|
||||
ssl_context = (
|
||||
ssl_verify
|
||||
if isinstance(ssl_verify, bool)
|
||||
else ssl.create_default_context(cafile=ssl_verify)
|
||||
)
|
||||
|
||||
async def send_request(session: aiohttp.ClientSession, dp_rank: int) -> int:
|
||||
json_data = {
|
||||
"sampling_params": {
|
||||
"temperature": 0.0,
|
||||
"max_new_tokens": 8,
|
||||
"ignore_eos": True,
|
||||
},
|
||||
"bootstrap_host": FAKE_BOOTSTRAP_HOST,
|
||||
"bootstrap_room": dp_rank,
|
||||
"input_ids": [10, 11, 12, 13],
|
||||
"routed_dp_rank": dp_rank,
|
||||
}
|
||||
async with session.post(
|
||||
url + "/generate", json=json_data, ssl=ssl_context
|
||||
) as response:
|
||||
await response.read()
|
||||
return response.status
|
||||
|
||||
async with aiohttp.ClientSession(
|
||||
timeout=aiohttp.ClientTimeout(total=timeout),
|
||||
headers=headers,
|
||||
) as session:
|
||||
return await asyncio.gather(
|
||||
*(send_request(session, dp_rank) for dp_rank in range(server_args.dp_size))
|
||||
)
|
||||
|
||||
|
||||
def _execute_server_warmup(server_args: ServerArgs):
|
||||
headers = {}
|
||||
url = server_args.url()
|
||||
@@ -2124,42 +2166,28 @@ def _execute_server_warmup(server_args: ServerArgs):
|
||||
|
||||
else:
|
||||
logger.info(f"Start of pd disaggregation warmup ...")
|
||||
request_name = "/generate"
|
||||
json_data = {
|
||||
"sampling_params": {
|
||||
"temperature": 0.0,
|
||||
"max_new_tokens": 8,
|
||||
"ignore_eos": True,
|
||||
},
|
||||
"bootstrap_host": [FAKE_BOOTSTRAP_HOST] * server_args.dp_size,
|
||||
# This is a hack to ensure fake transfer is enabled during prefill warmup
|
||||
# ensure each dp rank has a unique bootstrap_room during prefill warmup
|
||||
"bootstrap_room": [
|
||||
i * (2**63 // server_args.dp_size) + (i % server_args.tp_size)
|
||||
for i in range(server_args.dp_size)
|
||||
],
|
||||
"input_ids": [[10, 11, 12, 13]] * server_args.dp_size,
|
||||
}
|
||||
res = requests.post(
|
||||
url + request_name,
|
||||
json=json_data,
|
||||
headers=headers,
|
||||
timeout=(
|
||||
warmup_timeout if warmup_timeout > 0 else 1800
|
||||
), # because of deep gemm precache is very long if not precache.
|
||||
verify=ssl_verify,
|
||||
status_codes = asyncio.run(
|
||||
_send_disaggregation_warmup_requests(
|
||||
server_args=server_args,
|
||||
url=url,
|
||||
headers=headers,
|
||||
ssl_verify=ssl_verify,
|
||||
timeout=warmup_timeout if warmup_timeout > 0 else 1800,
|
||||
)
|
||||
)
|
||||
if res.status_code == 200:
|
||||
failed_status_codes = [code for code in status_codes if code != 200]
|
||||
if not failed_status_codes:
|
||||
logger.info(
|
||||
f"Disaggregation warmup request completed with status {res.status_code}, resp: {res.json()}"
|
||||
"Disaggregation warmup requests completed for all %s DP ranks",
|
||||
server_args.dp_size,
|
||||
)
|
||||
logger.info("End of disaggregation warmup")
|
||||
_global_state.tokenizer_manager.server_status = ServerStatus.Up
|
||||
else:
|
||||
logger.info(
|
||||
"Disaggregation warmup failed (mode=%s), status code: %s",
|
||||
"Disaggregation warmup failed (mode=%s), status codes: %s",
|
||||
server_args.disaggregation_mode,
|
||||
res.status_code,
|
||||
failed_status_codes,
|
||||
)
|
||||
_global_state.tokenizer_manager.server_status = ServerStatus.UnHealthy
|
||||
|
||||
@@ -2169,8 +2197,6 @@ def _execute_server_warmup(server_args: ServerArgs):
|
||||
kill_process_tree(os.getpid())
|
||||
return False
|
||||
|
||||
# Debug print
|
||||
# logger.info(f"warmup request returns: {res.json()=}")
|
||||
return success
|
||||
|
||||
|
||||
|
||||
@@ -0,0 +1,83 @@
|
||||
import asyncio
|
||||
import unittest
|
||||
from types import SimpleNamespace
|
||||
from unittest.mock import patch
|
||||
|
||||
from sglang.srt.disaggregation.utils import FAKE_BOOTSTRAP_HOST
|
||||
from sglang.srt.entrypoints.http_server import (
|
||||
_send_disaggregation_warmup_requests,
|
||||
)
|
||||
from sglang.test.ci.ci_register import register_cpu_ci
|
||||
|
||||
register_cpu_ci(est_time=5, suite="base-a-test-cpu")
|
||||
|
||||
|
||||
class TestDisaggregationServerWarmup(unittest.IsolatedAsyncioTestCase):
|
||||
async def test_sends_concurrent_scalar_request_to_each_dp_rank(self):
|
||||
server_args = SimpleNamespace(dp_size=4)
|
||||
all_started = asyncio.Event()
|
||||
calls = []
|
||||
sessions = []
|
||||
|
||||
class Response:
|
||||
status = 200
|
||||
|
||||
async def __aenter__(self):
|
||||
if len(calls) == server_args.dp_size:
|
||||
all_started.set()
|
||||
await asyncio.wait_for(all_started.wait(), timeout=5)
|
||||
return self
|
||||
|
||||
async def __aexit__(self, *args):
|
||||
pass
|
||||
|
||||
async def read(self):
|
||||
return b""
|
||||
|
||||
class Session:
|
||||
def __init__(self, **kwargs):
|
||||
self.kwargs = kwargs
|
||||
sessions.append(self)
|
||||
|
||||
async def __aenter__(self):
|
||||
return self
|
||||
|
||||
async def __aexit__(self, *args):
|
||||
pass
|
||||
|
||||
def post(self, *args, **kwargs):
|
||||
calls.append((args, kwargs))
|
||||
return Response()
|
||||
|
||||
with patch("sglang.srt.entrypoints.http_server.aiohttp.ClientSession", Session):
|
||||
status_codes = await _send_disaggregation_warmup_requests(
|
||||
server_args=server_args,
|
||||
url="http://localhost:30000",
|
||||
headers={"Authorization": "Bearer token"},
|
||||
ssl_verify=False,
|
||||
timeout=123,
|
||||
)
|
||||
|
||||
self.assertEqual(status_codes, [200] * server_args.dp_size)
|
||||
self.assertEqual(len(calls), server_args.dp_size)
|
||||
self.assertEqual(len(sessions), 1)
|
||||
self.assertEqual(
|
||||
sessions[0].kwargs["headers"], {"Authorization": "Bearer token"}
|
||||
)
|
||||
self.assertEqual(sessions[0].kwargs["timeout"].total, 123)
|
||||
|
||||
calls_by_rank = {
|
||||
kwargs["json"]["routed_dp_rank"]: (args, kwargs) for args, kwargs in calls
|
||||
}
|
||||
self.assertEqual(set(calls_by_rank), set(range(server_args.dp_size)))
|
||||
|
||||
for dp_rank, (args, kwargs) in calls_by_rank.items():
|
||||
self.assertEqual(args, ("http://localhost:30000/generate",))
|
||||
self.assertEqual(kwargs["json"]["input_ids"], [10, 11, 12, 13])
|
||||
self.assertEqual(kwargs["json"]["bootstrap_host"], FAKE_BOOTSTRAP_HOST)
|
||||
self.assertEqual(kwargs["json"]["bootstrap_room"], dp_rank)
|
||||
self.assertFalse(kwargs["ssl"])
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
Reference in New Issue
Block a user