Merge branch 'pytorch:main' into bucketized_avg_localSGD

docs/source/coordination.rst

@@ -0,0 +1,4 @@
+.. automodule:: torchft.coordination
+    :members:
+    :undoc-members:
+    :show-inheritance:

docs/source/index.rst

@@ -21,6 +21,7 @@ the entire training job.
     data
     checkpointing
     parameter_server
+    coordination
 
 
 License

pyproject.toml

@@ -33,9 +33,10 @@ dev = [
 
 [tool.maturin]
 features = ["pyo3/extension-module"]
+module-name = "torchft._torchft"
 
 [project.scripts]
-torchft_lighthouse = "torchft.torchft:lighthouse_main"
+torchft_lighthouse = "torchft._torchft:lighthouse_main"
 
 [tool.isort]
 multi_line_output = 3

src/lib.rs

@@ -30,15 +30,30 @@ use crate::torchftpb::manager_service_client::ManagerServiceClient;
 use crate::torchftpb::{CheckpointMetadataRequest, ManagerQuorumRequest, ShouldCommitRequest};
 use pyo3::prelude::*;
 
+/// ManagerServer is a GRPC server for the manager service.
+/// There should be one manager server per replica group (typically running on
+/// the rank 0 host). The individual ranks within a replica group should use
+/// ManagerClient to communicate with the manager server and participate in
+/// quorum operations.
+///
+/// Args:
+///     replica_id (str): The ID of the replica group.
+///     lighthouse_addr (str): The HTTP address of the lighthouse server.
+///     hostname (str): The hostname of the manager server.
+///     bind (str): The HTTP address to bind the server to.
+///     store_addr (str): The HTTP address of the store server.
+///     world_size (int): The world size of the replica group.
+///     heartbeat_interval (timedelta): The interval at which heartbeats are sent.
+///     connect_timeout (timedelta): The timeout for connecting to the lighthouse server.
 #[pyclass]
-struct Manager {
+struct ManagerServer {
     handle: JoinHandle<Result<()>>,
     manager: Arc<manager::Manager>,
     _runtime: Runtime,
 }
 
 #[pymethods]
-impl Manager {
+impl ManagerServer {
     #[new]
     fn new(
         py: Python<'_>,
@@ -74,17 +89,29 @@ impl Manager {
         })
     }
 
+    /// address returns the address of the manager server.
+    ///
+    /// Returns:
+    ///   str: The address of the manager server.
     fn address(&self) -> PyResult<String> {
         Ok(self.manager.address().to_string())
     }
 
+    /// shutdown shuts down the manager server.
     fn shutdown(&self, py: Python<'_>) {
         py.allow_threads(move || {
             self.handle.abort();
         })
     }
 }
 
+/// ManagerClient is a GRPC client to the manager service.
+///
+/// It is used by the trainer to communicate with the ManagerServer.
+///
+/// Args:
+///     addr (str): The HTTP address of the manager server.
+///     connect_timeout (timedelta): The timeout for connecting to the manager server.
 #[pyclass]
 struct ManagerClient {
     runtime: Runtime,
@@ -108,7 +135,7 @@ impl ManagerClient {
         })
     }
 
-    fn quorum(
+    fn _quorum(
         &self,
         py: Python<'_>,
         rank: i64,
@@ -147,7 +174,7 @@ impl ManagerClient {
         })
     }
 
-    fn checkpoint_metadata(
+    fn _checkpoint_metadata(
         &self,
         py: Python<'_>,
         rank: i64,
@@ -168,6 +195,20 @@ impl ManagerClient {
         })
     }
 
+    /// should_commit makes a request to the manager to determine if the trainer
+    /// should commit the current step. This waits until all ranks check in at
+    /// the specified step and will return false if any worker passes
+    /// ``should_commit=False``.
+    ///
+    /// Args:
+    ///     rank (int): The rank of the trainer.
+    ///     step (int): The step of the trainer.
+    ///     should_commit (bool): Whether the trainer should commit the current step.
+    ///     timeout (timedelta): The timeout for the request. If the request
+    ///         times out a TimeoutError is raised.
+    ///
+    /// Returns:
+    ///    bool: Whether the trainer should commit the current step.
     fn should_commit(
         &self,
         py: Python<'_>,
@@ -263,15 +304,28 @@ async fn lighthouse_main_async(opt: lighthouse::LighthouseOpt) -> Result<()> {
     Ok(())
 }
 
+/// LighthouseServer is a GRPC server for the lighthouse service.
+///
+/// It is used to coordinate the ManagerServer for each replica group.
+///
+/// This entrypoint is primarily for testing and debugging purposes. The
+/// ``torchft_lighthouse`` command is recommended for most use cases.
+///
+/// Args:
+///     bind (str): The HTTP address to bind the server to.
+///     min_replicas (int): The minimum number of replicas required to form a quorum.
+///     join_timeout_ms (int): The timeout for joining the quorum.
+///     quorum_tick_ms (int): The interval at which the quorum is checked.
+///     heartbeat_timeout_ms (int): The timeout for heartbeats.
 #[pyclass]
-struct Lighthouse {
+struct LighthouseServer {
     lighthouse: Arc<lighthouse::Lighthouse>,
     handle: JoinHandle<Result<()>>,
     _runtime: Runtime,
 }
 
 #[pymethods]
-impl Lighthouse {
+impl LighthouseServer {
     #[pyo3(signature = (bind, min_replicas, join_timeout_ms=None, quorum_tick_ms=None, heartbeat_timeout_ms=None))]
     #[new]
     fn new(
@@ -307,10 +361,15 @@ impl Lighthouse {
         })
     }
 
+    /// address returns the address of the lighthouse server.
+    ///
+    /// Returns:
+    ///    str: The address of the lighthouse server.
     fn address(&self) -> PyResult<String> {
         Ok(self.lighthouse.address().to_string())
     }
 
+    /// shutdown shuts down the lighthouse server.
     fn shutdown(&self, py: Python<'_>) {
         py.allow_threads(move || {
             self.handle.abort();
@@ -339,7 +398,7 @@ impl From<Status> for StatusError {
 }
 
 #[pymodule]
-fn torchft(m: &Bound<'_, PyModule>) -> PyResult<()> {
+fn _torchft(m: &Bound<'_, PyModule>) -> PyResult<()> {
     // setup logging on import
     let mut log = stderrlog::new();
     log.verbosity(2)
@@ -353,9 +412,9 @@ fn torchft(m: &Bound<'_, PyModule>) -> PyResult<()> {
     log.init()
         .map_err(|e| PyRuntimeError::new_err(e.to_string()))?;
 
-    m.add_class::<Manager>()?;
+    m.add_class::<ManagerServer>()?;
     m.add_class::<ManagerClient>()?;
-    m.add_class::<Lighthouse>()?;
+    m.add_class::<LighthouseServer>()?;
     m.add_class::<QuorumResult>()?;
     m.add_function(wrap_pyfunction!(lighthouse_main, m)?)?;
 

torchft/torchft.pyi → torchft/_torchft.pyi

@@ -3,15 +3,15 @@ from typing import List, Optional
 
 class ManagerClient:
     def __init__(self, addr: str, connect_timeout: timedelta) -> None: ...
-    def quorum(
+    def _quorum(
         self,
         rank: int,
         step: int,
         checkpoint_metadata: str,
         shrink_only: bool,
         timeout: timedelta,
     ) -> QuorumResult: ...
-    def checkpoint_metadata(self, rank: int, timeout: timedelta) -> str: ...
+    def _checkpoint_metadata(self, rank: int, timeout: timedelta) -> str: ...
     def should_commit(
         self,
         rank: int,
@@ -33,7 +33,7 @@ class QuorumResult:
     max_world_size: int
     heal: bool
 
-class Manager:
+class ManagerServer:
     def __init__(
         self,
         replica_id: str,
@@ -48,7 +48,7 @@ class Manager:
     def address(self) -> str: ...
     def shutdown(self) -> None: ...
 
-class Lighthouse:
+class LighthouseServer:
     def __init__(
         self,
         bind: str,

torchft/coordination.py

@@ -0,0 +1,23 @@
+"""
+Coordination (Low Level API)
+============================
+
+.. warning::
+    As torchft is still in development, the APIs in this module are subject to change.
+
+This module exposes low level coordination APIs to allow you to build your own 
+custom fault tolerance algorithms on top of torchft.
+
+If you're looking for a more complete solution, please use the other modules in
+torchft.
+
+This provides direct access to the Lighthouse and Manager servers and clients.
+"""
+
+from torchft._torchft import LighthouseServer, ManagerClient, ManagerServer
+
+__all__ = [
+    "LighthouseServer",
+    "ManagerServer",
+    "ManagerClient",
+]

torchft/coordination_test.py

@@ -0,0 +1,19 @@
+import inspect
+from unittest import TestCase
+
+from torchft.coordination import LighthouseServer, ManagerClient, ManagerServer
+
+
+class TestCoordination(TestCase):
+    def test_coordination_docs(self) -> None:
+        classes = [
+            ManagerClient,
+            ManagerServer,
+            LighthouseServer,
+        ]
+        for cls in classes:
+            self.assertIn("Args:", str(cls.__doc__), cls)
+            for name, method in inspect.getmembers(cls, predicate=inspect.ismethod):
+                if name.startswith("_"):
+                    continue
+                self.assertIn("Args:", str(cls.__doc__), cls)

torchft/lighthouse_test.py

@@ -4,15 +4,15 @@
 import torch.distributed as dist
 
 from torchft import Manager, ProcessGroupGloo
-from torchft.torchft import Lighthouse
+from torchft._torchft import LighthouseServer
 
 
 class TestLighthouse(TestCase):
     def test_join_timeout_behavior(self) -> None:
         """Test that join_timeout_ms affects joining behavior"""
         # To test, we create a lighthouse with 100ms and 400ms join timeouts
         # and measure the time taken to validate the quorum.
-        lighthouse = Lighthouse(
+        lighthouse = LighthouseServer(
             bind="[::]:0",
             min_replicas=1,
             join_timeout_ms=100,
@@ -52,14 +52,14 @@ def test_join_timeout_behavior(self) -> None:
             if "manager" in locals():
                 manager.shutdown()
 
-        lighthouse = Lighthouse(
+        lighthouse = LighthouseServer(
             bind="[::]:0",
             min_replicas=1,
             join_timeout_ms=400,
         )
 
     def test_heartbeat_timeout_ms_sanity(self) -> None:
-        lighthouse = Lighthouse(
+        lighthouse = LighthouseServer(
             bind="[::]:0",
             min_replicas=1,
             heartbeat_timeout_ms=100,

torchft/local_sgd_integ_test.py

@@ -8,11 +8,11 @@
 import torch
 from torch import nn, optim
 
+from torchft._torchft import LighthouseServer
 from torchft.local_sgd import DiLoCo, LocalSGD
 from torchft.manager import Manager
 from torchft.manager_integ_test import FailureInjector, MyModel, Runner
 from torchft.process_group import ProcessGroupGloo
-from torchft.torchft import Lighthouse
 
 logger: logging.Logger = logging.getLogger(__name__)
 
@@ -166,7 +166,7 @@ def state_dict() -> Dict[str, Dict[str, object]]:  # pyre-ignore[53]
 
 class ManagerIntegTest(TestCase):
     def test_local_sgd_recovery(self) -> None:
-        lighthouse = Lighthouse(
+        lighthouse = LighthouseServer(
             bind="[::]:0",
             min_replicas=2,
         )
@@ -214,7 +214,7 @@ def test_local_sgd_recovery(self) -> None:
         self.assertEqual(failure_injectors[1].count, 1)
 
     def test_diloco_healthy(self) -> None:
-        lighthouse = Lighthouse(
+        lighthouse = LighthouseServer(
             bind="[::]:0",
             min_replicas=2,
         )
@@ -258,7 +258,7 @@ def test_diloco_healthy(self) -> None:
                 )
 
     def test_diloco_recovery(self) -> None:
-        lighthouse = Lighthouse(
+        lighthouse = LighthouseServer(
             bind="[::]:0",
             min_replicas=2,
         )

torchft/manager.py

@@ -38,9 +38,9 @@
 import torch
 from torch.distributed import ReduceOp, TCPStore
 
+from torchft._torchft import ManagerClient, ManagerServer
 from torchft.checkpointing import CheckpointTransport, HTTPTransport
 from torchft.futures import future_timeout
-from torchft.torchft import Manager as _Manager, ManagerClient
 
 if TYPE_CHECKING:
     from torchft.process_group import ProcessGroup
@@ -180,7 +180,7 @@ def __init__(
             wait_for_workers=False,
         )
         self._pg = pg
-        self._manager: Optional[_Manager] = None
+        self._manager: Optional[ManagerServer] = None
 
         if rank == 0:
             if port is None:
@@ -192,7 +192,7 @@ def __init__(
             if replica_id is None:
                 replica_id = ""
             replica_id = replica_id + str(uuid.uuid4())
-            self._manager = _Manager(
+            self._manager = ManagerServer(
                 replica_id=replica_id,
                 lighthouse_addr=lighthouse_addr,
                 hostname=hostname,
@@ -429,7 +429,7 @@ def wait_quorum(self) -> None:
     def _async_quorum(
         self, allow_heal: bool, shrink_only: bool, quorum_timeout: timedelta
     ) -> None:
-        quorum = self._client.quorum(
+        quorum = self._client._quorum(
             rank=self._rank,
             step=self._step,
             checkpoint_metadata=self._checkpoint_transport.metadata(),
@@ -498,7 +498,7 @@ def _async_quorum(
                 primary_client = ManagerClient(
                     recover_src_manager_address, connect_timeout=self._connect_timeout
                 )
-                checkpoint_metadata = primary_client.checkpoint_metadata(
+                checkpoint_metadata = primary_client._checkpoint_metadata(
                     self._rank, timeout=self._timeout
                 )
                 recover_src_rank = quorum.recover_src_rank