torchmonarch-nightly 2025.7.1__cp311-cp311-manylinux2014_x86_64.whl → 2025.7.25__cp311-cp311-manylinux2014_x86_64.whl

tests/test_allocator.py

@@ -6,14 +6,17 @@
 
 # pyre-strict
 
+import asyncio
 import contextlib
 import importlib.resources
+import logging
 import math
 import os
 import subprocess
 import sys
 import unittest
 from datetime import timedelta
+from time import sleep
 from typing import Generator, Optional
 from unittest import mock
 
@@ -24,22 +27,29 @@ import torch
 import torch.distributed as dist
 import torch.nn.functional as F
 
-from monarch._rust_bindings.hyperactor_extension.alloc import (
-    AllocConstraints,
-    AllocSpec,
-)
+from monarch._rust_bindings.monarch_hyperactor.alloc import AllocConstraints, AllocSpec
 from monarch._rust_bindings.monarch_hyperactor.channel import (
     ChannelAddr,
     ChannelTransport,
 )
-from monarch.actor_mesh import Actor, current_rank, current_size, endpoint, ValueMesh
-from monarch.allocator import (
+
+from monarch._src.actor.actor_mesh import MonarchContext
+from monarch._src.actor.allocator import (
     ALLOC_LABEL_PROC_MESH_NAME,
+    LocalAllocator,
     RemoteAllocator,
     StaticRemoteAllocInitializer,
     TorchXRemoteAllocInitializer,
 )
-from monarch.proc_mesh import ProcMesh
+from monarch._src.actor.sync_state import fake_sync_state
+from monarch.actor import (
+    Actor,
+    current_rank,
+    current_size,
+    endpoint,
+    ProcMesh,
+    ValueMesh,
+)
 from monarch.tools.mesh_spec import MeshSpec, ServerSpec
 from monarch.tools.network import get_sockaddr
 
@@ -49,6 +59,19 @@ from torchx.specs import AppState
 _100_MILLISECONDS = timedelta(milliseconds=100)
 
 SERVER_READY = "monarch.tools.commands.server_ready"
+UNUSED = "__UNUSED__"
+
+
+class EnvCheckActor(Actor):
+    """Actor that checks for the presence of an environment variable"""
+
+    def __init__(self) -> None:
+        pass
+
+    @endpoint
+    async def get_env_var(self, var_name: str) -> str:
+        """Return the value of the specified environment variable or 'NOT_SET' if not found"""
+        return os.environ.get(var_name, "NOT_SET")
 
 
 class TestActor(Actor):
@@ -57,6 +80,8 @@ class TestActor(Actor):
     def __init__(self) -> None:
         self.rank: int = current_rank().rank
         self.world_size: int = math.prod(current_size().values())
+        self.logger: logging.Logger = logging.getLogger("test_actor")
+        self.logger.setLevel(logging.INFO)
 
     @endpoint
     async def compute_world_size(self, master_addr: str, master_port: int) -> int:
@@ -71,17 +96,33 @@ class TestActor(Actor):
         finally:
             dist.destroy_process_group()
 
+    @endpoint
+    async def log(self, message: str) -> None:
+        print(f"Stdout LogMessage from print: {message}")
+        sys.stderr.write(f"Stderr LogMessage from print: {message}\n")
+        self.logger.info(f"LogMessage from logger: {message}")
+
 
 @contextlib.contextmanager
-def remote_process_allocator(addr: Optional[str] = None) -> Generator[str, None, None]:
-    with importlib.resources.path(__package__, "") as package_path:
+def remote_process_allocator(
+    addr: Optional[str] = None, timeout: Optional[int] = None
+) -> Generator[str, None, None]:
+    """Start a remote process allocator on addr. If timeout is not None, have it
+    timeout after that many seconds if no messages come in"""
+
+    with importlib.resources.as_file(
+        importlib.resources.files(__package__)
+    ) as package_path:
         addr = addr or ChannelAddr.any(ChannelTransport.Unix)
+        args = [
+            "process_allocator",
+            f"--addr={addr}",
+        ]
+        if timeout is not None:
+            args.append(f"--timeout-sec={timeout}")
 
         process_allocator = subprocess.Popen(
-            args=[
-                "process_allocator",
-                f"--addr={addr}",
-            ],
+            args=args,
             env={
                 # prefix PATH with this test module's directory to
                 # give 'process_allocator' and 'monarch_bootstrap' binary resources
@@ -102,6 +143,82 @@ def remote_process_allocator(addr: Optional[str] = None) -> Generator[str, None,
                 process_allocator.kill()
 
 
+class TestSetupActorInAllocator(unittest.IsolatedAsyncioTestCase):
+    @classmethod
+    def setUpClass(cls) -> None:
+        cloudpickle.register_pickle_by_value(sys.modules[TestActor.__module__])
+
+    @classmethod
+    def tearDownClass(cls) -> None:
+        cloudpickle.unregister_pickle_by_value(sys.modules[TestActor.__module__])
+
+    async def test_setup_lambda_with_multiple_env_vars(self) -> None:
+        """Test that the setup lambda can set multiple environment variables"""
+        env_vars: dict[str, str] = {
+            "TEST_ENV_VAR_1": "value_1",
+            "TEST_ENV_VAR_2": "value_2",
+            "TEST_ENV_VAR_3": "value_3",
+        }
+
+        def setup_multiple_env_vars(ctx: MonarchContext) -> None:
+            for name, value in env_vars.items():
+                os.environ[name] = value
+
+        spec = AllocSpec(AllocConstraints(), gpus=1, hosts=1)
+        allocator = LocalAllocator()
+        alloc = await allocator.allocate(spec)
+
+        proc_mesh = await ProcMesh.from_alloc(alloc, setup=setup_multiple_env_vars)
+
+        try:
+            actor = await proc_mesh.spawn("env_check", EnvCheckActor)
+
+            for name, expected_value in env_vars.items():
+                actual_value = await actor.get_env_var.call_one(name)
+                self.assertEqual(
+                    actual_value,
+                    expected_value,
+                    f"Environment variable {name} was not set correctly",
+                )
+        finally:
+            await proc_mesh.stop()
+
+    async def test_setup_lambda_with_context_info(self) -> None:
+        """Test that the setup lambda can access context information"""
+        context_var_name: str = "PROC_MESH_CONTEXT_INFO"
+
+        def setup_with_context(ctx: MonarchContext) -> None:
+            context_info = f"proc_id:{ctx.proc_id},point_rank:{ctx.point.rank}"
+            os.environ[context_var_name] = context_info
+
+        spec = AllocSpec(AllocConstraints(), gpus=1, hosts=1)
+        allocator = LocalAllocator()
+        alloc = await allocator.allocate(spec)
+
+        proc_mesh = await ProcMesh.from_alloc(alloc, setup=setup_with_context)
+
+        try:
+            actor = await proc_mesh.spawn("env_check", EnvCheckActor)
+
+            context_info = await actor.get_env_var.call_one(context_var_name)
+
+            self.assertNotEqual(
+                context_info,
+                "NOT_SET",
+                "Context information was not stored in the environment variable",
+            )
+            self.assertIn(
+                "proc_id:", context_info, "Context information does not contain proc_id"
+            )
+            self.assertIn(
+                "point_rank:0",
+                context_info,
+                f"Context information {context_info} does not contain point_rank",
+            )
+        finally:
+            await proc_mesh.stop()
+
+
 class TestRemoteAllocator(unittest.IsolatedAsyncioTestCase):
     @classmethod
     def setUpClass(cls) -> None:
@@ -153,7 +270,7 @@ class TestRemoteAllocator(unittest.IsolatedAsyncioTestCase):
             """test initializer that returns an empty list of addresses"""
 
             async def initialize_alloc(self, match_labels: dict[str, str]) -> list[str]:
-                _ = match_labels  # Suppress unused variable warning
+                _ = match_labels
                 return []
 
         empty_initializer = EmptyAllocInitializer()
@@ -191,6 +308,209 @@ class TestRemoteAllocator(unittest.IsolatedAsyncioTestCase):
 
             self.assert_computed_world_size(values, world_size)
 
+    async def test_stop_proc_mesh_blocking(self) -> None:
+        spec = AllocSpec(AllocConstraints(), host=2, gpu=4)
+        with remote_process_allocator() as host1, remote_process_allocator() as host2:
+            allocator = RemoteAllocator(
+                world_id="test_remote_allocator",
+                initializer=StaticRemoteAllocInitializer(host1, host2),
+                heartbeat_interval=_100_MILLISECONDS,
+            )
+
+            alloc = await allocator.allocate(spec)
+            proc_mesh = await ProcMesh.from_alloc(alloc)
+            # XXX - it is not clear why this trying to use
+            # async code in a sync context.
+            with fake_sync_state():
+                actor = proc_mesh.spawn("test_actor", TestActor).get()
+                proc_mesh.stop().get()
+            with self.assertRaises(
+                RuntimeError, msg="`ProcMesh` has already been stopped"
+            ):
+                proc_mesh.spawn("test_actor", TestActor).get()
+            del actor
+
+    async def test_wrong_address(self) -> None:
+        hosts = 1
+        gpus = 1
+        spec = AllocSpec(AllocConstraints(), host=hosts, gpu=gpus)
+
+        # create 2x process-allocators (on their own bind addresses) to simulate 2 hosts
+        with remote_process_allocator():
+            wrong_host = ChannelAddr.any(ChannelTransport.Unix)
+            allocator = RemoteAllocator(
+                world_id="test_remote_allocator",
+                initializer=StaticRemoteAllocInitializer(wrong_host),
+                heartbeat_interval=_100_MILLISECONDS,
+            )
+            alloc = await allocator.allocate(spec)
+
+            with self.assertRaisesRegex(
+                Exception, r"no process has ever been allocated.*"
+            ):
+                await ProcMesh.from_alloc(alloc)
+
+    async def test_init_failure(self) -> None:
+        class FailInitActor(Actor):
+            def __init__(self) -> None:
+                if current_rank().rank == 0:
+                    raise RuntimeError("fail on init")
+
+            @endpoint
+            def dummy(self) -> None:
+                pass
+
+        with remote_process_allocator() as host1, remote_process_allocator() as host2:
+            allocator = RemoteAllocator(
+                world_id="helloworld",
+                initializer=StaticRemoteAllocInitializer(host1, host2),
+                heartbeat_interval=_100_MILLISECONDS,
+            )
+            spec = AllocSpec(AllocConstraints(), host=2, gpu=2)
+            proc_mesh = await ProcMesh.from_alloc(await allocator.allocate(spec))
+            actor_mesh = await proc_mesh.spawn("actor", FailInitActor)
+
+            with self.assertRaisesRegex(
+                Exception,
+                r"(?s)fail on init",
+            ):
+                await actor_mesh.dummy.call()
+
+    async def test_stop_proc_mesh(self) -> None:
+        spec = AllocSpec(AllocConstraints(), host=2, gpu=4)
+
+        # create 2x process-allocators (on their own bind addresses) to simulate 2 hosts
+        with remote_process_allocator() as host1, remote_process_allocator() as host2:
+            allocator = RemoteAllocator(
+                world_id="test_remote_allocator",
+                initializer=StaticRemoteAllocInitializer(host1, host2),
+                heartbeat_interval=_100_MILLISECONDS,
+            )
+            alloc = await allocator.allocate(spec)
+            proc_mesh = await ProcMesh.from_alloc(alloc)
+            actor = await proc_mesh.spawn("test_actor", TestActor)
+
+            await proc_mesh.stop()
+
+            with self.assertRaises(
+                RuntimeError, msg="`ProcMesh` has already been stopped"
+            ):
+                await proc_mesh.spawn("test_actor", TestActor)
+
+            # TODO(agallagher): It'd be nice to test that this just fails
+            # immediately, trying to access the wrapped actor mesh, but right
+            # now we doing casting without accessing the wrapped type.
+            del actor
+
+    async def test_stop_proc_mesh_context_manager(self) -> None:
+        spec = AllocSpec(AllocConstraints(), host=2, gpu=4)
+
+        # create 2x process-allocators (on their own bind addresses) to simulate 2 hosts
+        with remote_process_allocator() as host1, remote_process_allocator() as host2:
+            allocator = RemoteAllocator(
+                world_id="test_remote_allocator",
+                initializer=StaticRemoteAllocInitializer(host1, host2),
+                heartbeat_interval=_100_MILLISECONDS,
+            )
+            alloc = await allocator.allocate(spec)
+            proc_mesh = await ProcMesh.from_alloc(alloc)
+            with self.assertRaises(ValueError, msg="foo"):
+                async with proc_mesh:
+                    actor = await proc_mesh.spawn("test_actor", TestActor)
+                    # Ensure that proc mesh is stopped when context manager exits.
+                    raise ValueError("foo")
+
+            with self.assertRaises(
+                RuntimeError, msg="`ProcMesh` has already been stopped"
+            ):
+                await proc_mesh.spawn("test_actor", TestActor)
+
+            # TODO(agallagher): It'd be nice to test that this just fails
+            # immediately, trying to access the wrapped actor mesh, but right
+            # now we doing casting without accessing the wrapped type.
+            del actor
+
+    async def test_setup_lambda_sets_env_vars(self) -> None:
+        """Test that the setup lambda can set environment variables during proc_mesh allocation"""
+        test_var_name: str = "TEST_ENV_VAR_FOR_PROC_MESH"
+        test_var_value: str = "test_value_123"
+
+        def setup_env_vars(ctx: MonarchContext) -> None:
+            os.environ[test_var_name] = test_var_value
+
+        hosts = 2
+        gpus = 4
+        spec = AllocSpec(AllocConstraints(), host=hosts, gpu=gpus)
+
+        with remote_process_allocator() as host1, remote_process_allocator() as host2:
+            allocator = RemoteAllocator(
+                world_id="test_remote_allocator",
+                initializer=StaticRemoteAllocInitializer(host1, host2),
+                heartbeat_interval=_100_MILLISECONDS,
+            )
+            alloc = await allocator.allocate(spec)
+            proc_mesh = await ProcMesh.from_alloc(alloc, setup=setup_env_vars)
+
+            try:
+                actor = await proc_mesh.spawn("env_check", EnvCheckActor)
+
+                env_var_values = await actor.get_env_var.call(test_var_name)
+                env_var_value = env_var_values.item(host=0, gpu=0)
+
+                self.assertEqual(
+                    env_var_value,
+                    test_var_value,
+                    f"Environment variable {test_var_name} was not set correctly",
+                )
+            finally:
+                await proc_mesh.stop()
+
+    async def test_stop_proc_mesh_context_manager_multiple_times(self) -> None:
+        spec = AllocSpec(AllocConstraints(), host=2, gpu=4)
+
+        # create 2x process-allocators (on their own bind addresses) to simulate 2 hosts
+        with remote_process_allocator() as host1, remote_process_allocator() as host2:
+            allocator = RemoteAllocator(
+                world_id="test_remote_allocator",
+                initializer=StaticRemoteAllocInitializer(host1, host2),
+                heartbeat_interval=_100_MILLISECONDS,
+            )
+            alloc = await allocator.allocate(spec)
+            proc_mesh = await ProcMesh.from_alloc(alloc)
+            # We can nest multiple context managers on the same mesh, the innermost
+            # one closes the mesh and it cannot be used after that.
+            async with proc_mesh:
+                async with proc_mesh:
+                    actor = await proc_mesh.spawn("test_actor", TestActor)
+
+                with self.assertRaises(
+                    RuntimeError, msg="`ProcMesh` has already been stopped"
+                ):
+                    await proc_mesh.spawn("test_actor", TestActor)
+                # Exiting a second time should not raise an error.
+
+            # TODO(agallagher): It'd be nice to test that this just fails
+            # immediately, trying to access the wrapped actor mesh, but right
+            # now we doing casting without accessing the wrapped type.
+            del actor
+
+    async def test_remote_allocator_with_no_connection(self) -> None:
+        spec = AllocSpec(AllocConstraints(), host=1, gpu=4)
+
+        with remote_process_allocator(timeout=1) as host1:
+            # Wait 3 seconds without making any processes, make sure it dies.
+            await asyncio.sleep(3)
+            allocator = RemoteAllocator(
+                world_id="test_remote_allocator",
+                initializer=StaticRemoteAllocInitializer(host1),
+                heartbeat_interval=_100_MILLISECONDS,
+            )
+            with self.assertRaisesRegex(
+                Exception, "no process has ever been allocated on"
+            ):
+                alloc = await allocator.allocate(spec)
+                await ProcMesh.from_alloc(alloc)
+
     async def test_stacked_1d_meshes(self) -> None:
         # create two stacked actor meshes on the same host
         # each actor mesh running on separate process-allocators
@@ -244,7 +564,8 @@ class TestRemoteAllocator(unittest.IsolatedAsyncioTestCase):
         # but there are more than 1 mesh (hence ambiguous which mesh to allocate on)
 
         server = ServerSpec(
-            name="__UNUSED__",
+            name=UNUSED,
+            scheduler=UNUSED,
             state=AppState.RUNNING,
             meshes=[MeshSpec(name="x", num_hosts=1), MeshSpec(name="y", num_hosts=1)],
         )
@@ -262,7 +583,8 @@ class TestRemoteAllocator(unittest.IsolatedAsyncioTestCase):
     @pytest.mark.oss_skip  # pyre-ignore[56] TODO T228752279
     async def test_torchx_remote_alloc_initializer_no_match_label_1_mesh(self) -> None:
         server = ServerSpec(
-            name="__UNUSED__",
+            name=UNUSED,
+            scheduler=UNUSED,
             state=AppState.RUNNING,
             meshes=[
                 MeshSpec(
@@ -295,7 +617,8 @@ class TestRemoteAllocator(unittest.IsolatedAsyncioTestCase):
     @pytest.mark.oss_skip  # pyre-ignore[56] TODO T228752279
     async def test_torchx_remote_alloc_initializer_with_match_label(self) -> None:
         server = ServerSpec(
-            name="__UNUSED__",
+            name=UNUSED,
+            scheduler=UNUSED,
             state=AppState.RUNNING,
             meshes=[
                 MeshSpec(
@@ -338,6 +661,7 @@ class TestRemoteAllocator(unittest.IsolatedAsyncioTestCase):
 
         server = ServerSpec(
             name="test",
+            scheduler=UNUSED,
             state=AppState.RUNNING,
             meshes=[
                 MeshSpec(
@@ -363,3 +687,35 @@ class TestRemoteAllocator(unittest.IsolatedAsyncioTestCase):
                     )
                 )
                 await ProcMesh.from_alloc(alloc)
+
+    async def test_log(self) -> None:
+        # create a mesh to log to both stdout and stderr
+
+        with remote_process_allocator() as host:
+            allocator = RemoteAllocator(
+                world_id="test_actor_logger",
+                initializer=StaticRemoteAllocInitializer(host),
+                heartbeat_interval=_100_MILLISECONDS,
+            )
+
+            spec = AllocSpec(AllocConstraints(), host=1, gpu=2)
+
+            proc_mesh = await ProcMesh.from_alloc(await allocator.allocate(spec))
+
+            # Generate aggregated log every 1 second.
+            await proc_mesh.logging_option(True, 1)
+            actor = await proc_mesh.spawn("actor", TestActor)
+            # Run for 4 seconds, every second generates 5 logs, so we expect to see
+            # 2 actors x 5 logs/actor/sec * 1 sec = 10 logs per aggregation.
+            for _ in range(20):
+                await actor.log.call("Expect to see [10 processes]")
+                sleep(0.2)
+            # Generate aggregated log every 2 seconds.
+            await proc_mesh.logging_option(True, 2)
+            # Run for 8 seconds, every second generates 5 logs, so we expect to see
+            # 2 actors x 5 logs/actor/sec * 2 sec = 20 logs per aggregation.
+            for _ in range(40):
+                await actor.log.call("Expect to see [20 processes]")
+                sleep(0.2)
+
+            print("======== All Done ========")

tests/test_controller.py

@@ -653,6 +653,8 @@ def test_panicking_worker():
             _ = fetch_shard(torch.ones(2, 3)).result()
 
 
+# TODO - re-enable after resolving T232206970
+@pytest.mark.oss_skip
 def test_timeout_warning(caplog):
     timeout = 3
     with local_rust_device_mesh(