Compatibility for xgboost>=1.7.0, fix master CI (#242)

* Update requirements-test.txt

Signed-off-by: Antoni Baum <[email protected]>

* Update requirements-test.txt

Signed-off-by: Antoni Baum <[email protected]>

* Update requirements-test.txt

Signed-off-by: Antoni Baum <[email protected]>

* Update requirements-test.txt

Signed-off-by: Antoni Baum <[email protected]>

* Compatibility for xgboost 1.7.0

Signed-off-by: Antoni Baum <[email protected]>

* Fix MRO

Signed-off-by: Antoni Baum <[email protected]>

Signed-off-by: Antoni Baum <[email protected]>

Author: Yard1

requirements-test.txt

@@ -6,10 +6,11 @@ packaging
 petastorm
 pytest
 pyarrow
-ray[tune]
+ray[tune, data]
 scikit-learn
 modin
 dask
 
 #workaround for now
+protobuf<4.0.0
 tensorboardX==2.2

xgboost_ray/main.py

@@ -25,6 +25,17 @@
     class EarlyStopException(XGBoostError):
         pass
 
+
+# From xgboost>=1.7.0, rabit is replaced by a collective communicator
+try:
+    from xgboost.collective import CommunicatorContext
+    rabit = None
+    HAS_COLLECTIVE = True
+except ImportError:
+    from xgboost import rabit  # noqa
+    CommunicatorContext = None
+    HAS_COLLECTIVE = False
+
 from xgboost_ray.callback import DistributedCallback, \
     DistributedCallbackContainer
 from xgboost_ray.compat import TrainingCallback, RabitTracker, LEGACY_CALLBACK
@@ -66,7 +77,7 @@ def inner_f(*args, **kwargs):
     RayDeviceQuantileDMatrix, RayDataIter, concat_dataframes, \
     LEGACY_MATRIX
 from xgboost_ray.session import init_session, put_queue, \
-    set_session_queue
+    set_session_queue, get_rabit_rank
 
 
 def _get_environ(item: str, old_val: Any):
@@ -237,25 +248,40 @@ def _stop_rabit_tracker(rabit_process: multiprocessing.Process):
     rabit_process.terminate()
 
 
-class _RabitContext:
+class _RabitContextBase:
     """This context is used by local training actors to connect to the
     Rabit tracker.
 
     Args:
         actor_id (str): Unique actor ID
-        args (list): Arguments for Rabit initialisation. These are
+        args (dict): Arguments for Rabit initialisation. These are
             environment variables to configure Rabit clients.
     """
 
-    def __init__(self, actor_id, args):
+    def __init__(self, actor_id: int, args: dict):
+        args["DMLC_TASK_ID"] = "[xgboost.ray]:" + actor_id
         self.args = args
-        self.args.append(("DMLC_TASK_ID=[xgboost.ray]:" + actor_id).encode())
 
-    def __enter__(self):
-        xgb.rabit.init(self.args)
 
-    def __exit__(self, *args):
-        xgb.rabit.finalize()
+# From xgboost>=1.7.0, rabit is replaced by a collective communicator
+if HAS_COLLECTIVE:
+
+    class _RabitContext(_RabitContextBase, CommunicatorContext):
+        pass
+
+else:
+
+    class _RabitContext(_RabitContextBase):
+        def __init__(self, actor_id: int, args: dict):
+            super().__init__(actor_id, args)
+            self._list_args = [("%s=%s" % item).encode()
+                               for item in self.args.items()]
+
+        def __enter__(self):
+            xgb.rabit.init(self._list_args)
+
+        def __exit__(self, *args):
+            xgb.rabit.finalize()
 
 
 def _ray_get_actor_cpus():
@@ -517,12 +543,12 @@ def _save_checkpoint_callback(self):
 
         class _SaveInternalCheckpointCallback(TrainingCallback):
             def after_iteration(self, model, epoch, evals_log):
-                if xgb.rabit.get_rank() == 0 and \
+                if get_rabit_rank() == 0 and \
                         epoch % this.checkpoint_frequency == 0:
                     put_queue(_Checkpoint(epoch, pickle.dumps(model)))
 
             def after_training(self, model):
-                if xgb.rabit.get_rank() == 0:
+                if get_rabit_rank() == 0:
                     put_queue(_Checkpoint(-1, pickle.dumps(model)))
                 return model
 
@@ -1054,8 +1080,7 @@ def handle_actor_failure(actor_id):
     maybe_log("[RayXGBoost] Starting XGBoost training.")
 
     # Start Rabit tracker for gradient sharing
-    rabit_process, env = _start_rabit_tracker(alive_actors)
-    rabit_args = [("%s=%s" % item).encode() for item in env.items()]
+    rabit_process, rabit_args = _start_rabit_tracker(alive_actors)
 
     # Load checkpoint if we have one. In that case we need to adjust the
     # number of training rounds.

xgboost_ray/session.py

@@ -63,7 +63,11 @@ def get_actor_rank() -> int:
 @PublicAPI
 def get_rabit_rank() -> int:
     import xgboost as xgb
-    return xgb.rabit.get_rank()
+    try:
+        # From xgboost>=1.7.0, rabit is replaced by a collective communicator
+        return xgb.collective.get_rank()
+    except (ImportError, AttributeError):
+        return xgb.rabit.get_rank()
 
 
 @PublicAPI