torchmonarch-nightly 2025.8.2__cp313-cp313-manylinux2014_x86_64.whl → 2025.9.3__cp313-cp313-manylinux2014_x86_64.whl

monarch/tools/config/workspace.py

@@ -0,0 +1,165 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# pyre-strict
+
+import shutil
+from pathlib import Path
+
+from monarch.tools.config.environment import CondaEnvironment, Environment
+
+ACTIVE_CONDA_ENV = CondaEnvironment()
+
+
+class Workspace:
+    """
+    A workspace is one or more local directories that contains your project(s).
+    Workspaces can specify an "environment" on which projects are developed and run locally.
+    A currently active conda environment is an example of such environment.
+
+    At the time of job submission an ephemeral version of the "image" is built and the
+    new job is configured to run on this image. The "image" is the one specified by
+    `Role.image` attribute in the job's `AppDef`
+    (see `monarch.tools.components.hyperactor.host_mesh()`).
+
+    For example when launching onto Kubernetes, "image" is interpreted as a Docker image (e.g. "name:tag")
+
+    Specifically the ephemeral image contains:
+
+    1. A copy of the workspace directories
+    2. (If Applicable) A copy of the currently active environment
+
+    This effectively one-time mirrors the local codebase and environment on the remote machines.
+
+    Workspaces can also be sync'ed interactively on-demand (post job launch) by using
+    `monarch.actor.proc_mesh.ProcMesh.syncWorkspace(Workspace)`.
+
+    Usage:
+
+    .. doc-test::
+
+        import pathlib
+        from monarch.tools.config import Workspace
+        from monarch.tools.config import Config
+
+        HOME = pathlib.Path().home()
+
+        # 1. single project workspace
+        config = Config(
+            workspace=Workspace(dirs=[HOME / "github" / "torchtitan"]),
+        )
+
+        # 2. multiple projects (useful for cross-project development)
+        config = Config(
+            workspace=Workspace(
+                dirs=[
+                    # $HOME/torch             (local) -> $WORKSPACE_DIR/torch      (remote)
+                    # $HOME/github/torchtitan (local) -> $WORKSPACE_DIR/torchtitan (remote)
+                    HOME() / "torch",
+                    HOME() / "github" / "torchtitan",
+                ]
+            ),
+        )
+
+        # 3. with explicit local -> remote mappings
+        config = Config(
+            workspace=Workspace(
+                dirs={
+                    # $HOME/torch             (local) -> $WORKSPACE_DIR/github/pytorch    (remote)
+                    # $HOME/github/torchtitan (local) -> $WORKSPACE_DIR/github/torchtitan (remote)
+                    HOME() / "torch" : "github/pytorch"
+                    HOME() / "github" / "torchtitan" : "github/torchtitan"
+                }
+            )
+        )
+        # -- or flat into WORKSPACE_DIR
+        config = Config(
+            workspace=Workspace(
+                # $HOME/github/torchtitan  (local) -> $WORKSPACE_DIR/  (remote)
+                dirs={HOME() / "github" / "torchtitan": ""},
+            )
+        )
+
+        # 3. no project, everything is installed in my environment (but sync my env)
+        config = Config(
+            workspace=Workspace(),
+        )
+
+        # 4. disable project and environment sync
+        config = Config(
+            workspace=Workspace(env=None),
+        )
+    """
+
+    def __init__(
+        self,
+        dirs: list[Path | str] | dict[Path | str, str] | None = None,
+        env: Environment | None = ACTIVE_CONDA_ENV,
+    ) -> None:
+        self.env = env
+        self.dirs: dict[Path, str] = {}  # src -> dst
+
+        if dirs is None:
+            pass
+        elif isinstance(dirs, list):
+            for d in dirs:
+                d = Path(d)
+                self.dirs[d] = d.name
+        else:  # dict
+            for src, dst in dirs.items():
+                self.dirs[Path(src)] = dst
+
+    def __eq__(self, other: object) -> bool:
+        if not isinstance(other, Workspace):
+            return False
+
+        return self.env == other.env and self.dirs == other.dirs
+
+    def merge(self, outdir: str | Path) -> None:
+        """Merges the dirs of this workspace into the given outdir."""
+
+        outdir = Path(outdir)
+        outdir.mkdir(parents=True, exist_ok=True)
+
+        for src, dst in self.dirs.items():
+            shutil.copytree(src, outdir / dst, dirs_exist_ok=True)
+
+    # pyre-ignore[2] skip type-hint to avoid torchx dep
+    def set_env_vars(self, appdef) -> None:
+        """For each role in the appdef, sets the following env vars (if not already set):
+
+        1. `WORKSPACE_DIR`: the root directory of the remote workspace
+        2. `PYTHONPATH`: include all the remote workspace dirs for all the roles in the appdef
+                (dedups and appends to existing `PYTHONPATH`)
+        3. `CONDA_DIR`: (if env is conda) the remote path to the conda env to activate
+        """
+
+        # typically this macro comes from torchx.specs.macros.img_root
+        # but we use the str repr instead to avoid taking a dep to torchx from this module
+        # unittest (test_workspace.py) asserts against torchx.specs.macros.img_root
+        # guarding against changes to the macro value
+        img_root_macro = "${img_root}"
+
+        for role in appdef.roles:
+            remote_workspace_root = role.env.setdefault(
+                "WORKSPACE_DIR",
+                f"{img_root_macro}/workspace",
+            )
+
+            PYTHONPATH = [p for p in role.env.get("PYTHONPATH", "").split(":") if p]
+            for dst in self.dirs.values():
+                remote_dir = f"{remote_workspace_root}/{dst}"
+                if remote_dir not in PYTHONPATH:
+                    PYTHONPATH.append(remote_dir)
+            role.env["PYTHONPATH"] = ":".join(PYTHONPATH)
+
+            if isinstance(self.env, CondaEnvironment):
+                role.env.setdefault("CONDA_DIR", f"{img_root_macro}/conda")
+
+    @staticmethod
+    def null() -> "Workspace":
+        """Returns a "null" workspace; a workspace with no project dirs and no environment."""
+        return Workspace(env=None)

monarch/tools/mesh_spec.py

@@ -128,6 +128,7 @@ class ServerSpec:
     meshes: list[MeshSpec]
     scheduler: str
     namespace: str = ""
+    ui_url: Optional[str] = None
 
     @property
     def server_handle(self) -> str:
@@ -210,6 +211,7 @@ class ServerSpec:
         return {
             "name": self.name,
             "server_handle": self.server_handle,
+            **({"ui_url": self.ui_url} if self.ui_url else {}),
             "state": self.state.name,
             "meshes": {
                 mesh.name: {

monarch/utils/__init__.py

@@ -0,0 +1,9 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .utils import setup_env_for_distributed
+
+__all__ = ["setup_env_for_distributed"]

monarch/utils/utils.py

@@ -0,0 +1,78 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+# pyre-strict
+import os
+import socket
+
+from monarch.actor import Actor, current_rank, endpoint, ProcMesh
+
+
+def _find_free_port() -> int:
+    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
+        s.bind(("localhost", 0))
+        addr = s.getsockname()
+        port = addr[1]
+        return port
+
+
+class _TorchDistributedInitActor(Actor):
+    def __init__(self) -> None:
+        self.rank: int = current_rank().rank
+
+    @endpoint
+    def get_host_port(self) -> tuple[str, int]:
+        return (socket.gethostname(), _find_free_port())
+
+    @endpoint
+    def setup_env(self, master_addr: str, master_port: int) -> None:
+        cr = current_rank()
+        # Assume last dimension is the local rank.
+        last_label = cr.extent.labels[-1]
+        local_world_size = cr.size(last_label)
+        world_size = cr.extent.nelements
+        global_rank = cr.rank
+        local_rank = min(world_size, global_rank % local_world_size)
+        group_rank = global_rank // local_world_size
+        group_world_size = (world_size + local_world_size - 1) // local_world_size
+        env = {
+            "MASTER_ADDR": master_addr,
+            "MASTER_PORT": str(master_port),
+            "RANK": str(global_rank),
+            "LOCAL_RANK": str(local_rank),
+            "LOCAL_WORLD_SIZE": str(local_world_size),
+            "GROUP_RANK": str(group_rank),
+            "GROUP_WORLD_SIZE": str(group_world_size),
+            "ROLE_RANK": str(global_rank),
+            "ROLE_WORLD_SIZE": str(world_size),
+            "ROLE_NAME": "rank",
+            "WORLD_SIZE": str(world_size),
+        }
+        os.environ.update(env)
+
+
+async def setup_env_for_distributed(
+    proc_mesh: ProcMesh,
+    master_addr: str | None = None,
+    master_port: int | None = None,
+) -> None:
+    """
+    Sets up environment variables for pytorch distributed.
+    It selects a random proc in the proc_mesh to be the master node.
+    It sets enviornment variables like RANK, LOCAL_RANK, WORLD_SIZE, etc.
+    If master_addr and master_port are None, it will automatically select a master node and port.
+    """
+    assert (
+        (master_addr is None) == (master_port is None)
+    ), "Either both master_addr and master_port must be specified or neither must be specified."
+    am = await proc_mesh.spawn("_TorchDistributedInitActor", _TorchDistributedInitActor)
+    if master_addr is None:
+        # We use call instead of call_one because call_one can't handle tuple return types.
+        vm = await am.flatten("rank").slice(rank=0).get_host_port.call()
+        master_addr, master_port = vm.item()
+    assert master_port is not None, "master_port should not be None here."
+    await am.setup_env.call(master_addr, master_port)

tests/error_test_binary.py

@@ -4,9 +4,10 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+# pyre-unsafe
+
 import asyncio
 import ctypes
-import sys
 
 import click
 from monarch._rust_bindings.monarch_extension.blocking import blocking_function
@@ -158,8 +159,9 @@ def error_endpoint(num_procs, sync_test_impl, sync_endpoint, endpoint_name):
 @main.command("error-bootstrap")
 def error_bootstrap():
     print("Started function error_bootstrap", flush=True)
-
-    proc_mesh(gpus=4, env={"MONARCH_ERROR_DURING_BOOTSTRAP_FOR_TESTING": "1"}).get()
+    proc_mesh(
+        gpus=4, env={"MONARCH_ERROR_DURING_BOOTSTRAP_FOR_TESTING": "1"}
+    ).initialized.get()
 
 
 async def _error_unmonitored():

tests/python_actor_test_binary.py

@@ -0,0 +1,52 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# pyre-strict
+
+import asyncio
+import logging
+
+import click
+
+from monarch.actor import Actor, endpoint, proc_mesh
+
+
+@click.group()
+def main() -> None:
+    pass
+
+
+class Printer(Actor):
+    def __init__(self) -> None:
+        self.logger: logging.Logger = logging.getLogger()
+
+    @endpoint
+    async def print(self, content: str) -> None:
+        print(f"{content}", flush=True)
+
+
+async def _flush_logs() -> None:
+    # Create a lot of processes to stress test the logging
+    pm = await proc_mesh(gpus=32)
+
+    # never flush
+    await pm.logging_option(aggregate_window_sec=1000)
+    am = await pm.spawn("printer", Printer)
+
+    # These should be streamed to client
+    for _ in range(5):
+        await am.print.call("has print streaming")
+
+    await pm.stop()
+
+
+@main.command("flush-logs")
+def flush_logs() -> None:
+    asyncio.run(_flush_logs())
+
+
+if __name__ == "__main__":
+    main()

tests/test_actor_error.py

@@ -4,6 +4,7 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+# pyre-unsafe
 
 import importlib.resources
 import os
@@ -34,6 +35,24 @@ class ExceptionActorSync(Actor):
         raise Exception("This is a test exception")
 
 
+class NestedExceptionActor(Actor):
+    @endpoint
+    async def raise_exception_with_context(self) -> None:
+        try:
+            raise Exception("Inner exception")
+        except Exception:
+            # Don't use from here to set __context__ instead of __cause__
+            raise Exception("Outer exception")
+
+    @endpoint
+    async def raise_exception_with_cause(self) -> None:
+        try:
+            raise Exception("Inner exception")
+        except Exception as e:
+            # Use from here to set __cause__ instead of __context__
+            raise Exception("Outer exception") from e
+
+
 class BrokenPickleClass:
     """A class that can be configured to raise exceptions during pickling/unpickling."""
 
@@ -116,6 +135,41 @@ def test_actor_exception_sync(mesh, actor_class, num_procs):
             exception_actor.raise_exception.call().get()
 
 
+@pytest.mark.parametrize(
+    "mesh",
+    [local_proc_mesh, proc_mesh],
+    ids=["local_proc_mesh", "distributed_proc_mesh"],
+)
+async def test_actor_error_message(mesh):
+    """
+    Test that exceptions raised in actor endpoints capture nested exceptions.
+    """
+    proc = mesh(gpus=2)
+    exception_actor = await proc.spawn("exception_actor", NestedExceptionActor)
+
+    with pytest.raises(ActorError) as exc_info:
+        await exception_actor.raise_exception_with_cause.call()
+
+    # Make sure both exception messages are present in the message.
+    assert "Inner exception" in str(exc_info.value)
+    assert "Outer exception" in str(exc_info.value)
+    # Make sure the "cause" is set.
+    assert "The above exception was the direct cause of the following exception" in str(
+        exc_info.value
+    )
+
+    with pytest.raises(ActorError) as exc_info:
+        await exception_actor.raise_exception_with_context.call()
+
+    # Make sure both exception messages are present in the message.
+    assert "Inner exception" in str(exc_info.value)
+    assert "Outer exception" in str(exc_info.value)
+    # Make sure the "cause" is set.
+    assert "During handling of the above exception, another exception occurred" in str(
+        exc_info.value
+    )
+
+
 '''
 # oss_skip: importlib not pulling resource correctly in git CI, needs to be revisited
 @pytest.mark.oss_skip
@@ -436,14 +490,14 @@ async def test_proc_mesh_monitoring(mesh):
     event = await anext(monitor)
     assert isinstance(event, ProcEvent.Crashed)
     assert event[0] == 0  # check rank
-    assert "ActorFailureError" in event[1]  # check error message
+    assert "failed: did not handle supervision event" in event[1]  # check error message
     assert (
         "Simulated actor failure for supervision testing" in event[1]
     )  # check error message
 
     # should not be able to spawn actors anymore as proc mesh is unhealthy
     with pytest.raises(SupervisionError, match="proc mesh is stopped with reason"):
-        await proc.spawn("ex", ExceptionActorSync)
+        await proc.spawn("ex", ExceptionActorSync).initialized
 
 
 @pytest.mark.parametrize(
@@ -467,16 +521,19 @@ async def test_actor_mesh_supervision_handling(mesh):
     await e.check.call()
 
     # existing call should fail with supervision error
-    with pytest.raises(SupervisionError, match="supervision error:"):
+    with pytest.raises(
+        SupervisionError,
+        match=".*Actor .* exited because of the following reason",
+    ):
         await e.fail_with_supervision_error.call_one()
 
     # new call should fail with check of health state of actor mesh
-    with pytest.raises(SupervisionError, match="actor mesh is not in a healthy state"):
+    with pytest.raises(SupervisionError, match="Actor .* is unhealthy with reason"):
         await e.check.call()
 
     # should not be able to spawn actors anymore as proc mesh is unhealthy
     with pytest.raises(SupervisionError, match="proc mesh is stopped with reason"):
-        await proc.spawn("ex", ExceptionActorSync)
+        await proc.spawn("ex", ExceptionActorSync).initialized
 
 
 class HealthyActor(Actor):
@@ -534,11 +591,14 @@ async def test_actor_mesh_supervision_handling_chained_error(mesh):
     # in a chain of client -> Intermediate -> ErrorActor, a supervision error
     # happening in ErrorActor will be captured by Intermediate and re-raised
     # as an application error (ActorError).
-    with pytest.raises(ActorError, match="supervision error:"):
+    with pytest.raises(
+        ActorError,
+        match=".*Actor .* exited because of the following reason",
+    ):
         await intermediate_actor.forward_error.call()
 
     # calling success endpoint should fail with ActorError, but with supervision msg.
-    with pytest.raises(ActorError, match="actor mesh is not in a healthy state"):
+    with pytest.raises(ActorError, match="Actor .* is unhealthy with reason"):
         await intermediate_actor.forward_success.call()
 
     # healthy actor should still be working
@@ -567,11 +627,14 @@ async def test_base_exception_handling(mesh, method_name):
     method = getattr(error_actor, method_name)
 
     # The call should raise a SupervisionError
-    with pytest.raises(SupervisionError, match="supervision error:"):
+    with pytest.raises(
+        SupervisionError,
+        match=".*Actor .* exited because of the following reason",
+    ):
         await method.call_one()
 
     # Subsequent calls should fail with a health state error
-    with pytest.raises(SupervisionError, match="actor mesh is not in a healthy state"):
+    with pytest.raises(RuntimeError, match="Actor .* is unhealthy with reason"):
         await error_actor.check.call()
 
 
@@ -587,18 +650,24 @@ async def test_supervision_with_proc_mesh_stopped(mesh):
     await proc.stop()
 
     # new call should fail with check of health state of actor mesh
-    with pytest.raises(SupervisionError, match="actor mesh is not in a healthy state"):
+    with pytest.raises(
+        SupervisionError, match="actor mesh is stopped due to proc mesh shutdown"
+    ):
         await actor_mesh.check.call()
 
     # proc mesh cannot spawn new actors anymore
     with pytest.raises(RuntimeError, match="`ProcMesh` has already been stopped"):
-        await proc.spawn("immediate", Intermediate)
+        await proc.spawn("immediate", Intermediate).initialized
 
 
 # TODO - re-enable after resolving T232206970
 @pytest.mark.oss_skip
 async def test_supervision_with_sending_error():
+    # Messages of length > this will cause a send error and a returned
+    # undeliverable.
     os.environ["HYPERACTOR_CODEC_MAX_FRAME_LENGTH"] = "50000000"
+    # Limit retries for sending before giving up.
+    os.environ["HYPERACTOR_MESSAGE_DELIVERY_TIMEOUT_SECS"] = "5"
 
     proc = await proc_mesh(gpus=1)
     actor_mesh = await proc.spawn("healthy", HealthyActor)
@@ -610,12 +679,71 @@ async def test_supervision_with_sending_error():
 
     # send a large payload to trigger send timeout error
     with pytest.raises(
-        SupervisionError, match="supervision error:.*actor mesh is stopped"
+        SupervisionError,
+        match=".*Actor .* exited because of the following reason",
     ):
         await actor_mesh.check_with_payload.call(payload="a" * 55000000)
 
     # new call should fail with check of health state of actor mesh
-    with pytest.raises(SupervisionError, match="actor mesh is not in a healthy state"):
+    with pytest.raises(SupervisionError, match="Actor .* is unhealthy with reason"):
         await actor_mesh.check.call()
-    with pytest.raises(SupervisionError, match="actor mesh is not in a healthy state"):
+    with pytest.raises(SupervisionError, match="Actor .* is unhealthy with reason"):
         await actor_mesh.check_with_payload.call(payload="a")
+
+
+async def test_slice_supervision() -> None:
+    pm = await proc_mesh(gpus=4)
+    healthy_mesh = await pm.spawn("healthy", HealthyActor)
+    error_mesh = await pm.spawn("error", ErrorActor)
+    slice_1 = error_mesh.slice(gpus=slice(2, 4))
+    slice_2 = error_mesh.slice(gpus=2)
+    slice_3 = error_mesh.slice(gpus=3)
+
+    # Trigger supervision error on gpus=3
+    with pytest.raises(SupervisionError, match="did not handle supervision event"):
+        await slice_3.fail_with_supervision_error.call()
+
+    # Mesh containing all gpus is unhealthy
+    with pytest.raises(SupervisionError, match="Actor .* is unhealthy with reason:"):
+        await error_mesh.check.call()
+
+    # Slice containing only gpus=3 is unhealthy
+    with pytest.raises(SupervisionError, match="Actor .* is unhealthy with reason:"):
+        await slice_3.check.call()
+
+    # Slice containing gpus=3 is unhealthy
+    with pytest.raises(SupervisionError, match="Actor .* is unhealthy with reason:"):
+        await slice_1.check.call()
+
+    # Slice not containing gpus=3 is healthy
+    check = await slice_2.check.call()
+    for _, item in check.items():
+        assert item == "this is a healthy check"
+
+    # Other actor mesh on the same proc mesh is healthy
+    check = await healthy_mesh.check.call()
+    for _, item in check.items():
+        assert item == "this is a healthy check"
+
+
+async def test_mesh_slices_inherit_parent_errors() -> None:
+    pm = await proc_mesh(gpus=4)
+    error_mesh = await pm.spawn("error", ErrorActor)
+    slice_1 = error_mesh.slice(gpus=slice(2, 4))
+
+    # Trigger supervision error on gpus=2, 3, 4
+    with pytest.raises(SupervisionError):
+        await slice_1.fail_with_supervision_error.call()
+
+    # Newly created slice containing gpu=3 is unhealthy
+    slice_2 = error_mesh.slice(gpus=3)
+    with pytest.raises(SupervisionError):
+        await slice_2.check.call()
+
+    # Newly created slice containing gpu=1 is healthy
+    slice_3 = error_mesh.slice(gpus=1)
+    check = await slice_3.check.call()
+    for _, item in check.items():
+        assert item == "this is a healthy check"
+
+    await pm.stop()

tests/test_alloc.py

@@ -20,6 +20,6 @@ class TestAlloc(IsolatedAsyncioTestCase):
         cmd = "echo hello"
         allocator = ProcessAllocator(cmd)
         spec = AllocSpec(AllocConstraints(), replica=2)
-        alloc = await allocator.allocate(spec)
+        alloc = allocator.allocate(spec)
 
         print(alloc)