Federated Learning not working with 2.1.0

[gubertoli]

Hi,
I am running a code based on this FL test code: https://github.com/dmlc/xgboost/tree/master/tests/test_distributed/test_federated

Using XGBoost 2.0.0 with world_size=5:

[18:28:59] Insecure federated server listening on 0.0.0.0:9091, world size 5
[18:29:00] XGBoost federated mode detected, not splitting data among workers
[18:29:00] XGBoost federated mode detected, not splitting data among workers
[18:29:00] XGBoost federated mode detected, not splitting data among workers
[18:29:00] XGBoost federated mode detected, not splitting data among workers
[18:29:00] XGBoost federated mode detected, not splitting data among workers
[18:29:00] XGBoost federated mode detected, not splitting data among workers
[18:29:00] XGBoost federated mode detected, not splitting data among workers
[18:29:00] XGBoost federated mode detected, not splitting data among workers
[18:29:00] XGBoost federated mode detected, not splitting data among workers
[18:29:00] XGBoost federated mode detected, not splitting data among workers
[18:29:04] [0]  eval-logloss:0.22646    train-logloss:0.23316
[18:29:06] [1]  eval-logloss:0.13776    train-logloss:0.13654
[18:29:07] [2]  eval-logloss:0.08036    train-logloss:0.08243
[18:29:09] [3]  eval-logloss:0.05830    train-logloss:0.05645
[18:29:11] [4]  eval-logloss:0.03825    train-logloss:0.04148
[18:29:12] [5]  eval-logloss:0.02660    train-logloss:0.02958
[18:29:14] [6]  eval-logloss:0.01386    train-logloss:0.01918
[18:29:16] [7]  eval-logloss:0.01018    train-logloss:0.01331
[18:29:17] [8]  eval-logloss:0.00847    train-logloss:0.01112
[18:29:19] [9]  eval-logloss:0.00691    train-logloss:0.00662
[18:29:21] [10] eval-logloss:0.00543    train-logloss:0.00503
[18:29:23] [11] eval-logloss:0.00445    train-logloss:0.00420
[18:29:24] [12] eval-logloss:0.00336    train-logloss:0.00355
[18:29:26] [13] eval-logloss:0.00277    train-logloss:0.00280
[18:29:28] [14] eval-logloss:0.00252    train-logloss:0.00244
[18:29:30] [15] eval-logloss:0.00177    train-logloss:0.00193
[18:29:31] [16] eval-logloss:0.00156    train-logloss:0.00161
[18:29:33] [17] eval-logloss:0.00135    train-logloss:0.00142
[18:29:35] [18] eval-logloss:0.00123    train-logloss:0.00125
[18:29:37] [19] eval-logloss:0.00106    train-logloss:0.00107
[18:29:37] Finished training

⚠️ ❗ However, after upgrading to XGBoost 2.1.0, the same code results in the following:

[18:33:03] Insecure federated server listening on 0.0.0.0:5, world size 9091
E0628 18:33:03.598596287  331756 chttp2_server.cc:1053]      UNKNOWN:No address added out of total 1 resolved for '0.0.0.0:5' {file:"/grpc/src/core/ext/transport/chttp2/server/chttp2_server.cc", file_line:963, created_time:"2024-06-28T18:33:03.597969893+02:00", children:[UNKNOWN:Failed to add any wildcard listeners {created_time:"2024-06-28T18:33:03.597933892+02:00", file_line:363, file:"/grpc/src/core/lib/iomgr/tcp_server_posix.cc", children:[UNKNOWN:Unable to configure socket {fd:8, created_time:"2024-06-28T18:33:03.597880883+02:00", file_line:220, file:"/grpc/src/core/lib/iomgr/tcp_server_utils_posix_common.cc", children:[UNKNOWN:Permission denied {syscall:"bind", os_error:"Permission denied", errno:13, created_time:"2024-06-28T18:33:03.597840858+02:00", file_line:194, file:"/grpc/src/core/lib/iomgr/tcp_server_utils_posix_common.cc"}]}, UNKNOWN:Unable to configure socket {file:"/grpc/src/core/lib/iomgr/tcp_server_utils_posix_common.cc", file_line:220, created_time:"2024-06-28T18:33:03.597922845+02:00", fd:8, children:[UNKNOWN:Permission denied {file:"/grpc/src/core/lib/iomgr/tcp_server_utils_posix_common.cc", file_line:194, created_time:"2024-06-28T18:33:03.597916832+02:00", errno:13, os_error:"Permission denied", syscall:"bind"}]}]}]}
[18:33:04] Rank 0
Process Process-2:
Traceback (most recent call last):
  File "/usr/lib/python3.10/multiprocessing/process.py", line 314, in _bootstrap
    self.run()
  File "/usr/lib/python3.10/multiprocessing/process.py", line 108, in run
    self._target(*self._args, **self._kwargs)
  File ".../refs/test_federated.py", line 36, in run_worker
    with xgb.collective.CommunicatorContext(**communicator_env):
  File ".../.venv/lib/python3.10/site-packages/xgboost/collective.py", line 280, in __enter__
    assert is_distributed()
AssertionError
[18:33:04] Rank 0
Process Process-3:
Traceback (most recent call last):
  File "/usr/lib/python3.10/multiprocessing/process.py", line 314, in _bootstrap
    self.run()
  File "/usr/lib/python3.10/multiprocessing/process.py", line 108, in run
    self._target(*self._args, **self._kwargs)
  File ".../refs/test_federated.py", line 36, in run_worker
    with xgb.collective.CommunicatorContext(**communicator_env):
  File ".../.venv/lib/python3.10/site-packages/xgboost/collective.py", line 280, in __enter__
    assert is_distributed()
AssertionError
[18:33:04] Rank 0
Process Process-4:
Traceback (most recent call last):
  File "/usr/lib/python3.10/multiprocessing/process.py", line 314, in _bootstrap
    self.run()
  File "/usr/lib/python3.10/multiprocessing/process.py", line 108, in run
    self._target(*self._args, **self._kwargs)
  File ".../refs/test_federated.py", line 36, in run_worker
    with xgb.collective.CommunicatorContext(**communicator_env):
  File ".../.venv/lib/python3.10/site-packages/xgboost/collective.py", line 280, in __enter__
    assert is_distributed()
AssertionError
[18:33:04] Rank 0
Process Process-5:
Traceback (most recent call last):
  File "/usr/lib/python3.10/multiprocessing/process.py", line 314, in _bootstrap
    self.run()
  File "/usr/lib/python3.10/multiprocessing/process.py", line 108, in run
    self._target(*self._args, **self._kwargs)
  File ".../refs/test_federated.py", line 36, in run_worker
    with xgb.collective.CommunicatorContext(**communicator_env):
  File ".../.venv/lib/python3.10/site-packages/xgboost/collective.py", line 280, in __enter__
    assert is_distributed()
AssertionError
[18:33:04] Rank 0
Process Process-6:
Traceback (most recent call last):
  File "/usr/lib/python3.10/multiprocessing/process.py", line 314, in _bootstrap
    self.run()
  File "/usr/lib/python3.10/multiprocessing/process.py", line 108, in run
    self._target(*self._args, **self._kwargs)
  File ".../refs/test_federated.py", line 36, in run_worker
    with xgb.collective.CommunicatorContext(**communicator_env):
  File ".../.venv/lib/python3.10/site-packages/xgboost/collective.py", line 280, in __enter__
    assert is_distributed()
AssertionError

For reference, the adapted test_federated.py is:

#!/usr/bin/python
import multiprocessing
import sys
import time

import xgboost as xgb
import xgboost.federated

SERVER_KEY = 'server-key.pem'
SERVER_CERT = 'server-cert.pem'
CLIENT_KEY = 'client-key.pem'
CLIENT_CERT = 'client-cert.pem'


def run_server(port: int, world_size: int, with_ssl: bool) -> None:
    if with_ssl:
        xgboost.federated.run_federated_server(port, world_size, SERVER_KEY, SERVER_CERT,
                                               CLIENT_CERT)
    else:
        xgboost.federated.run_federated_server(port, world_size)


def run_worker(port: int, world_size: int, rank: int, with_ssl: bool, with_gpu: bool) -> None:
    communicator_env = {
        'xgboost_communicator': 'federated',
        'federated_server_address': f'localhost:{port}',
        'federated_world_size': world_size,
        'federated_rank': rank
    }
    if with_ssl:
        communicator_env['federated_server_cert'] = SERVER_CERT
        communicator_env['federated_client_key'] = CLIENT_KEY
        communicator_env['federated_client_cert'] = CLIENT_CERT

    # Always call this before using distributed module
    with xgb.collective.CommunicatorContext(**communicator_env):
        # Load file, file will not be sharded in federated mode.
        dtrain = xgb.DMatrix('agaricus.txt.train-%02d?format=libsvm' % rank)
        dtest = xgb.DMatrix('agaricus.txt.test-%02d?format=libsvm' % rank)

        # Specify parameters via map, definition are same as c++ version
        param = {'max_depth': 2, 'eta': 1, 'objective': 'binary:logistic'}
        if with_gpu:
            param['tree_method'] = 'hist'
            param['device'] = f"cuda:{rank}"

        # Specify validations set to watch performance
        watchlist = [(dtest, 'eval'), (dtrain, 'train')]
        num_round = 20

        # Run training, all the features in training API is available.
        bst = xgb.train(param, dtrain, num_round, evals=watchlist,
                        early_stopping_rounds=2)

        # Save the model, only ask process 0 to save the model.
        if xgb.collective.get_rank() == 0:
            bst.save_model("test.model.json")
            xgb.collective.communicator_print("Finished training\n")


def run_federated(with_ssl: bool = True, with_gpu: bool = False) -> None:
    port = 9091
    world_size = int(sys.argv[1])

    server = multiprocessing.Process(target=run_server, args=(port, world_size, with_ssl))
    server.start()
    time.sleep(1)
    if not server.is_alive():
        raise Exception("Error starting Federated Learning server")

    workers = []
    for rank in range(world_size):
        worker = multiprocessing.Process(target=run_worker,
                                         args=(port, world_size, rank, with_ssl, with_gpu))
        workers.append(worker)
        worker.start()
    for worker in workers:
        worker.join()
    server.terminate()


if __name__ == '__main__':
    run_federated(with_ssl=False, with_gpu=False)

And the adapted shell script:

#!/bin/bash

# world_size=$(nvidia-smi -L | wc -l)
world_size=$1

# Split train and test files manually to simulate a federated environment.
split -n l/"${world_size}" -d agaricus.txt.train agaricus.txt.train-
split -n l/"${world_size}" -d agaricus.txt.test agaricus.txt.test-

python test_federated.py "${world_size}"

🚧 ⌛ Interim solution: Downgrade to XGBoost 2.0.0

[gubertoli]

@trivialfis I tested with 2.1.0 and 2.1.1 and the issue persists: #10716