Fix nccl future execution (#126)

Author: H-Huang

torchft/manager_integ_test.py

@@ -2,24 +2,26 @@
 import logging
 import threading
 import time
+import traceback
 from concurrent.futures import ThreadPoolExecutor, as_completed
 from contextlib import ExitStack, contextmanager
 from dataclasses import dataclass, field
 from datetime import timedelta
-from typing import Any, Dict, Generator, List, Protocol, Set, Tuple
+from typing import Any, Dict, Generator, List, Optional, Protocol, Set, Tuple, TypeVar
 from unittest import TestCase
 
 import torch
 import torch.distributed as dist
 from parameterized import parameterized
 from torch import nn, optim
+from torch._dynamo.utils import timed
 
 from torchft._torchft import LighthouseServer
 from torchft.ddp import DistributedDataParallel
 from torchft.local_sgd import DiLoCo, LocalSGD
 from torchft.manager import Manager
 from torchft.optim import OptimizerWrapper
-from torchft.process_group import ProcessGroupGloo
+from torchft.process_group import ProcessGroupBabyNCCL, ProcessGroupGloo
 
 logger: logging.Logger = logging.getLogger(__name__)
 
@@ -69,10 +71,14 @@ def check(self, rank: int, step: int) -> None:
                 raise InjectedFailure(f"injected failure {rank=} {step=}")
 
 
-class TrainLoop(Protocol):
+# R for an arbitrary return type
+R = TypeVar("R", covariant=True)
+
+
+class TrainLoop(Protocol[R]):
     def __call__(
         self, rank: int, store_port: int, device: torch.device, runner: "Runner"
-    ) -> Dict[str, Dict[str, object]]: ...
+    ) -> R: ...
 
 
 @dataclass
@@ -81,15 +87,15 @@ class Runner:
     num_replicas: int
     lighthouse_address: str
     failure_injector: FailureInjector
-    train_loop: TrainLoop
+    train_loop: TrainLoop[object]
 
     use_cuda: bool = False
     world_size: int = 1
     attempts: int = 3
     manager_args: Dict[str, object] = field(default_factory=dict)
     train_loop_args: Dict[str, Any] = field(default_factory=dict)
 
-    def _replica_main(self) -> List[Dict[str, Dict[str, object]]]:
+    def _replica_main(self) -> List[object]:
         store = dist.TCPStore(
             host_name="localhost",
             port=0,
@@ -131,7 +137,7 @@ def _replica_main(self) -> List[Dict[str, Dict[str, object]]]:
 
             return [fut.result() for fut in futures]
 
-    def run_replica(self) -> List[Dict[str, Dict[str, object]]]:
+    def run_replica(self) -> List[object]:
         for i in range(self.attempts):
             try:
                 print(
@@ -391,3 +397,92 @@ def test_quorum_timeout(self) -> None:
                     "status: Cancelled, message.*Timeout expired",
                 ):
                     manager.should_commit(timeout=timedelta(seconds=0.01))
+
+    @parameterized.expand(
+        [
+            (True,),  # Test with CUDA
+            (False,),  # Test without CUDA (CPU)
+        ]
+    )
+    def test_manager_allreduce(self, use_cuda: bool) -> None:
+        # Skip the test if use_cuda is True and there are not enough GPUs
+        if use_cuda and torch.cuda.device_count() < 2:
+            self.skipTest("Not enough GPUs for CUDA test")
+
+        # manager supports allreduce but we found an issue where the future callback is getting called
+        # before the allreduce is complete. This test is to ensure that the callback has stream synchronization
+        lighthouse = LighthouseServer(
+            bind="[::]:0",
+            min_replicas=2,
+        )
+        num_replicas = 2
+        futures = []
+
+        with ThreadPoolExecutor(max_workers=num_replicas) as executor:
+            for replica_id in range(num_replicas):
+                failure_injector = FailureInjector()
+                runner = Runner(
+                    replica_id=replica_id,
+                    num_replicas=num_replicas,
+                    lighthouse_address=lighthouse.address(),
+                    failure_injector=failure_injector,
+                    train_loop=all_reduce_callback,
+                    use_cuda=use_cuda,
+                )
+                futures.append(executor.submit(runner.run_replica))
+
+        results = []
+        for fut in as_completed(futures):
+            try:
+                results.append(fut.result()[0])
+            except Exception as e:
+                print(e, flush=True)
+                traceback.print_exc()
+                raise
+
+        lighthouse.shutdown()
+
+        print(results)
+        r0, r1 = results
+        torch.testing.assert_close(r0, r1, check_device=False)
+
+
+def all_reduce_callback(
+    rank: int,
+    store_port: int,
+    device: torch.device,
+    runner: Runner,
+) -> Optional[torch.Tensor]:
+    with ExitStack() as stack:
+        print(f"worker {runner.replica_id=} {rank=} {runner.world_size=} starting")
+
+        if device.type == "cuda":
+            pg = ProcessGroupBabyNCCL()
+        else:
+            pg = ProcessGroupGloo()
+        manager = Manager(
+            pg=pg,
+            min_replica_size=2,
+            use_async_quorum=False,
+            load_state_dict=lambda x: None,
+            state_dict=lambda: None,
+            replica_id=str(runner.replica_id),
+            store_addr="localhost",
+            store_port=store_port,
+            rank=rank,
+            world_size=runner.world_size,
+            lighthouse_addr=runner.lighthouse_address,
+            port=19530 + runner.replica_id,
+            timeout=timedelta(seconds=10),
+            quorum_timeout=timedelta(seconds=10),
+            # pyre-fixme[6]: Incompatible parameter type
+            **runner.manager_args,
+        )
+        stack.callback(lambda: manager.shutdown(wait=False))
+
+        manager.start_quorum()
+        t1 = torch.ones((1, 3), device=device)
+        fut = manager.allreduce(t1)
+        fut.wait()
+        return t1
+    return None

torchft/process_group.py

@@ -1093,10 +1093,12 @@ def _worker(
 
                         args = _PickleSafeOptions.unsafe_args(args)
                         fn = getattr(pg, func_name)
+
                         work[op_id] = _OpMetadata(
                             work=fn(*args, **kwargs),
                             stream=stream,
                         )
+
                 elif cmd == "wait":
                     op_id, timeout = cast(tuple[int, timedelta], op[1:])
 
@@ -1126,15 +1128,29 @@ def _worker(
                     del work[op_id]
                 elif cmd == "future":
                     op_id: int = cast(int, op[1])
+                    metadata: _OpMetadata = work[op_id]
 
-                    def callback(fut: Future[object]) -> None:
+                    def callback(fut: Future[object], metadata: _OpMetadata) -> None:
                         try:
-                            fut.wait()
-                            future_pipe.send((op_id, _FUTURE_RESULT, None))
+                            # create an event after the collective has been issued
+                            # to wait on this before we call "future"
+                            with metadata.set_stream():
+                                fut.wait()
+                                event = (
+                                    torch.cuda.current_stream().record_event(
+                                        torch.cuda.Event(interprocess=True)
+                                    )
+                                    if metadata.stream is not None
+                                    else None
+                                )
+
+                            future_pipe.send((op_id, _FUTURE_RESULT, None, event))
                         except Exception as e:
-                            future_pipe.send((op_id, _FUTURE_EXCEPTION, e))
+                            future_pipe.send((op_id, _FUTURE_EXCEPTION, e, None))
 
-                    work[op_id].work.get_future().add_done_callback(callback)
+                    metadata.work.get_future().add_done_callback(
+                        lambda fut: callback(fut, metadata)
+                    )
                 elif cmd == "num_active_work":
                     req_pipe.send(len(work))
                 else:
@@ -1153,11 +1169,15 @@ def _future_handler(self, future_pipe: _MonitoredPipe) -> None:
                 except TimeoutError:
                     continue
 
-                op_id, mode, data = cast(Tuple[int, str, object], cmd)
+                op_id, mode, data, event = cast(
+                    Tuple[int, str, object, Optional[torch.cuda.Event]], cmd
+                )
                 with self._futures_lock:
                     fut = self._futures[op_id]
                     del self._futures[op_id]
                 if mode == _FUTURE_RESULT:
+                    if event is not None:
+                        event.wait()
                     fut.set_result(data)
                 elif mode == _FUTURE_EXCEPTION:
                     fut.set_exception(data)