torchmonarch-nightly 2025.7.25__cp311-cp311-manylinux2014_x86_64.whl → 2025.7.26__cp311-cp311-manylinux2014_x86_64.whl

monarch/_src/actor/actor_mesh.py

@@ -65,6 +65,7 @@ from monarch._src.actor.endpoint import (
     Endpoint,
     EndpointProperty,
     Extent,
+    NotAnEndpoint,
     Propagator,
     Selection,
 )
@@ -76,7 +77,7 @@ from monarch._src.actor.pickle import flatten, unflatten
 from monarch._src.actor.shape import MeshTrait, NDSlice
 from monarch._src.actor.sync_state import fake_sync_state
 
-from monarch._src.actor.tensor_engine_shim import actor_send
+from monarch._src.actor.tensor_engine_shim import actor_rref, actor_send
 
 if TYPE_CHECKING:
     from monarch._src.actor.proc_mesh import ProcMesh
@@ -313,8 +314,7 @@ class ActorEndpoint(Endpoint[P, R]):
         """
         self._signature.bind(None, *args, **kwargs)
         objects, bytes = flatten((args, kwargs), _is_ref_or_mailbox)
-        refs = [obj for obj in objects if hasattr(obj, "__monarch_ref__")]
-        if not refs:
+        if all(not hasattr(obj, "__monarch_ref__") for obj in objects):
             message = PythonMessage(
                 PythonMessageKind.CallMethod(
                     self._name, None if port is None else port._port_ref
@@ -323,7 +323,7 @@ class ActorEndpoint(Endpoint[P, R]):
             )
             self._actor_mesh.cast(message, selection)
         else:
-            actor_send(self, bytes, refs, port, selection)
+            actor_send(self, bytes, objects, port, selection)
         shape = self._actor_mesh._shape
         return Extent(shape.labels, shape.ndslice.sizes)
 
@@ -335,6 +335,26 @@ class ActorEndpoint(Endpoint[P, R]):
             ), "unexpected receiver type"
         return PortTuple(p, PortReceiver(self._mailbox, self._supervise(r._receiver)))
 
+    def _rref(self, args, kwargs):
+        self._signature.bind(None, *args, **kwargs)
+        refs, bytes = flatten((args, kwargs), _is_ref_or_mailbox)
+
+        return actor_rref(self, bytes, refs)
+
+
+def as_endpoint(
+    not_an_endpoint: Callable[P, R], *, propagate: Propagator = None
+) -> Endpoint[P, R]:
+    if not isinstance(not_an_endpoint, NotAnEndpoint):
+        raise ValueError("expected an method of a spawned actor")
+    return ActorEndpoint(
+        not_an_endpoint._ref._actor_mesh_ref,
+        not_an_endpoint._name,
+        getattr(not_an_endpoint._ref, not_an_endpoint._name),
+        not_an_endpoint._ref._mailbox,
+        propagate,
+    )
+
 
 class Accumulator(Generic[P, R, A]):
     def __init__(
@@ -625,18 +645,23 @@ class _Actor:
                         f" This is likely due to an earlier error: {self._saved_error}"
                     )
                 raise AssertionError(error_message)
-            the_method = getattr(self.instance, method)._method
+            the_method = getattr(self.instance, method)
+            if isinstance(the_method, EndpointProperty):
+                module = the_method._method.__module__
+                the_method = functools.partial(the_method._method, self.instance)
+            else:
+                module = the_method.__module__
 
             if inspect.iscoroutinefunction(the_method):
 
                 async def instrumented():
                     enter_span(
-                        the_method.__module__,
+                        module,
                         method,
                         str(ctx.mailbox.actor_id),
                     )
                     try:
-                        result = await the_method(self.instance, *args, **kwargs)
+                        result = await the_method(*args, **kwargs)
                         self._maybe_exit_debugger()
                     except Exception as e:
                         logging.critical(
@@ -649,9 +674,9 @@ class _Actor:
 
                 result = await instrumented()
             else:
-                enter_span(the_method.__module__, method, str(ctx.mailbox.actor_id))
+                enter_span(module, method, str(ctx.mailbox.actor_id))
                 with fake_sync_state():
-                    result = the_method(self.instance, *args, **kwargs)
+                    result = the_method(*args, **kwargs)
                 self._maybe_exit_debugger()
                 exit_span()
 
@@ -758,35 +783,14 @@ class ActorMeshRef(MeshTrait):
                         attr_name,
                         attr_value._method,
                         self._mailbox,
+                        attr_value._propagator,
                     ),
                 )
 
-    def __getattr__(self, name: str) -> Any:
-        # This method is called when an attribute is not found
-        # For linting purposes, we need to tell the type checker that any attribute
-        # could be an endpoint that's dynamically added at runtime
-        # At runtime, we still want to raise AttributeError for truly missing attributes
-
-        # Check if this is a method on the underlying class
-        if hasattr(self._class, name):
-            attr = getattr(self._class, name)
-            if isinstance(attr, EndpointProperty):
-                # Dynamically create the endpoint
-                endpoint = ActorEndpoint(
-                    self._actor_mesh_ref,
-                    name,
-                    attr._method,
-                    self._mailbox,
-                    propagator=attr._propagator,
-                )
-                # Cache it for future use
-                setattr(self, name, endpoint)
-                return endpoint
-
-        # If we get here, it's truly not found
-        raise AttributeError(
-            f"'{self.__class__.__name__}' object has no attribute '{name}'"
-        )
+    def __getattr__(self, attr: str) -> NotAnEndpoint:
+        if attr in dir(self._class):
+            return NotAnEndpoint(self, attr)
+        raise AttributeError(attr)
 
     def _create(
         self,

monarch/_src/actor/endpoint.py

@@ -34,6 +34,7 @@ from monarch._src.actor.tensor_engine_shim import _cached_propagation, fake_call
 
 if TYPE_CHECKING:
     from monarch._src.actor.actor_mesh import (
+        ActorMeshRef,
         HyPortReceiver,
         OncePortReceiver,
         Port,
@@ -182,11 +183,22 @@ class Endpoint(ABC, Generic[P, R]):
         # pyre-ignore
         send(self, args, kwargs)
 
+    @abstractmethod
+    def _rref(self, args, kwargs) -> Any: ...
+
+    def rref(self, *args: P.args, **kwargs: P.kwargs) -> R:
+        return self._rref(args, kwargs)
+
     def _propagate(self, args, kwargs, fake_args, fake_kwargs):
         if self._propagator_arg is None or self._propagator_arg == "cached":
             if self._cache is None:
                 self._cache = {}
-            return _cached_propagation(self._cache, self._resolvable, args, kwargs)
+            resolvable = getattr(self, "_resolvable", None)
+            if resolvable is None:
+                raise NotImplementedError(
+                    "Cached propagation is not implemented for actor endpoints."
+                )
+            return _cached_propagation(self._cache, resolvable, args, kwargs)
         elif self._propagator_arg == "inspect":
             return None
         elif self._propagator_arg == "mocked":
@@ -229,13 +241,34 @@ class EndpointProperty(Generic[P, R]):
         return cast(Endpoint[P, R], self)
 
 
+class NotAnEndpoint:
+    """
+    Used as the dynamic value of functions on an ActorMeshRef that were not marked as endpoints.
+    This is used both to give a better error message (since we cannot prevent the type system from thinking they are methods),
+    and to provide the oppurtunity for someone to do endpoint(x.foo) on something that wasn't marked as an endpoint.
+    """
+
+    def __init__(self, ref: "ActorMeshRef", name: str):
+        self._ref = ref
+        self._name = name
+
+    def __call__(self, *args, **kwargs) -> None:
+        raise RuntimeError(
+            f"Actor {self._ref._class}.{self._name} is not annotated as an endpoint. To call it as one, add a @endpoint decorator to it, or directly wrap it in one as_endpoint(obj.method).call(...)"
+        )
+
+
 # This can't just be Callable because otherwise we are not
 # allowed to use type arguments in the return value.
 class EndpointIfy:
     @overload
-    def __call__(self, function: Callable[P, Awaitable[R]]) -> Endpoint[P, R]: ...
+    def __call__(
+        self, function: Callable[Concatenate[Any, P], Awaitable[R]]
+    ) -> Endpoint[P, R]: ...
     @overload
-    def __call__(self, function: Callable[P, R]) -> Endpoint[P, R]: ...
+    def __call__(
+        self, function: Callable[Concatenate[Any, P], R]
+    ) -> Endpoint[P, R]: ...
 
     def __call__(self, function: Any):
         pass

monarch/_src/actor/event_loop.py

@@ -14,7 +14,7 @@ import logging
 import threading
 from typing import Optional
 
-from libfb.py.pyre import none_throws
+from pyre_extensions import none_throws
 
 logger = logging.getLogger(__name__)
 

monarch/_src/actor/proc_mesh.py

@@ -43,7 +43,6 @@ from monarch._src.actor.actor_mesh import (
     Actor,
     ActorMeshRef,
     fake_sync_state,
-    MonarchContext,
 )
 
 from monarch._src.actor.allocator import LocalAllocator, ProcessAllocator, SimAllocator
@@ -89,7 +88,7 @@ class SetupActor(Actor):
     Typically used to setup the environment variables.
     """
 
-    def __init__(self, env: Callable[[MonarchContext], None]) -> None:
+    def __init__(self, env: Callable[[], None]) -> None:
         """
         Initialize the setup actor with the user defined setup method.
         """
@@ -100,8 +99,7 @@ class SetupActor(Actor):
         """
         Call the user defined setup method with the monarch context.
         """
-        ctx = MonarchContext.get()
-        self._setup_method(ctx)
+        self._setup_method()
 
 
 T = TypeVar("T")
@@ -114,7 +112,7 @@ except ImportError:
 
 
 async def _allocate_nonblocking(
-    alloc: Alloc, setup: Callable[[MonarchContext], None] | None = None
+    alloc: Alloc, setup: Callable[[], None] | None = None
 ) -> "ProcMesh":
     _proc_mesh = await HyProcMesh.allocate_nonblocking(alloc)
     if setup is None:
@@ -211,7 +209,7 @@ class ProcMesh(MeshTrait):
 
     @classmethod
     def from_alloc(
-        self, alloc: Alloc, setup: Callable[[MonarchContext], None] | None = None
+        self, alloc: Alloc, setup: Callable[[], None] | None = None
     ) -> Future["ProcMesh"]:
         """
         Allocate a process mesh according to the provided alloc.
@@ -219,7 +217,17 @@ class ProcMesh(MeshTrait):
 
         Arguments:
         - `alloc`: The alloc to allocate according to.
-        - `setup`: A lambda taking MonarchContext as param, can be used to setup env vars on the allocated mesh
+        - `setup`: An optional lambda function to configure environment variables on the allocated mesh.
+        Use the `current_rank()` method within the lambda to obtain the rank.
+
+        Example of a setup method to initialize torch distributed environment variables:
+        ```
+        def setup():
+            rank = current_rank()
+            os.environ["RANK"] = str(rank)
+            os.environ["WORLD_SIZE"] = str(len(rank.shape))
+            os.environ["LOCAL_RANK"] = str(rank["gpus"])
+        ```
         """
         return Future(
             impl=lambda: _allocate_nonblocking(alloc, setup),
@@ -428,7 +436,7 @@ async def proc_mesh_nonblocking(
     gpus: Optional[int] = None,
     hosts: int = 1,
     env: dict[str, str] | None = None,
-    setup: Callable[[MonarchContext], None] | None = None,
+    setup: Callable[[], None] | None = None,
 ) -> ProcMesh:
     if gpus is None:
         gpus = _local_device_count()
@@ -457,7 +465,7 @@ def proc_mesh(
     gpus: Optional[int] = None,
     hosts: int = 1,
     env: dict[str, str] | None = None,
-    setup: Callable[[MonarchContext], None] | None = None,
+    setup: Callable[[], None] | None = None,
 ) -> Future[ProcMesh]:
     return Future(
         impl=lambda: proc_mesh_nonblocking(

monarch/_src/actor/tensor_engine_shim.py

@@ -19,7 +19,6 @@ time it is used.
 
 if TYPE_CHECKING:
     from monarch._src.actor.actor_mesh import ActorEndpoint, Port, Selection
-    from monarch._src.actor.endpoint import Endpoint
 
 
 def shim(fn=None, *, module=None):
@@ -48,8 +47,12 @@ def actor_send(
 ) -> None: ...
 
 
+@shim(module="monarch.mesh_controller")
+def actor_rref(endpoint, args_kwargs_tuple: bytes, refs: Sequence[Any]): ...
+
+
 @shim(module="monarch.common.remote")
-def _cached_propagation(_cache, rfunction: "Endpoint", args, kwargs) -> Any: ...
+def _cached_propagation(_cache, rfunction, args, kwargs) -> Any: ...
 
 
 @shim(module="monarch.common.fake")

monarch/actor/__init__.py

@@ -12,6 +12,7 @@ from monarch._src.actor.actor_mesh import (
     Accumulator,
     Actor,
     ActorError,
+    as_endpoint,
     current_actor_name,
     current_rank,
     current_size,
@@ -35,6 +36,7 @@ __all__ = [
     "Actor",
     "ActorError",
     "current_actor_name",
+    "as_endpoint",
     "current_rank",
     "current_size",
     "endpoint",

monarch/common/messages.py

@@ -435,6 +435,15 @@ class SendResultOfActorCall(NamedTuple):
     stream: tensor_worker.StreamRef
 
 
+class CallActorMethod(NamedTuple):
+    seq: int
+    result: object
+    broker_id: Tuple[str, int]
+    local_state: Sequence[Tensor | tensor_worker.Ref]
+    mutates: List[tensor_worker.Ref]
+    stream: tensor_worker.StreamRef
+
+
 class SplitComm(NamedTuple):
     dims: Dims
     device_mesh: DeviceMesh

monarch/common/remote.py

@@ -157,7 +157,7 @@ class Remote(Generic[P, R], Endpoint[P, R]):
     def _maybe_resolvable(self):
         return None if self._remote_impl is None else self._resolvable
 
-    def rref(self, *args: P.args, **kwargs: P.kwargs) -> R:
+    def _rref(self, args, kwargs):
         return dtensor_dispatch(
             self._resolvable,
             self._propagate,
@@ -352,7 +352,7 @@ _miss = 0
 _hit = 0
 
 
-def _cached_propagation(_cache, rfunction: Endpoint, args, kwargs):
+def _cached_propagation(_cache, rfunction: ResolvableFunction, args, kwargs):
     tensors, shape_key = hashable_tensor_flatten(args, kwargs)
     # pyre-ignore
     inputs_group = TensorGroup([t._fake for t in tensors])

monarch/mesh_controller.py

@@ -30,6 +30,7 @@ from monarch._rust_bindings.monarch_extension.client import (  # @manual=//monar
     WorldState,
 )
 from monarch._rust_bindings.monarch_extension.mesh_controller import _Controller
+from monarch._rust_bindings.monarch_extension.tensor_worker import Ref
 from monarch._rust_bindings.monarch_hyperactor.actor import (
     PythonMessage,
     PythonMessageKind,
@@ -44,10 +45,12 @@ from monarch._src.actor.endpoint import Selection
 from monarch._src.actor.shape import NDSlice
 from monarch.common import device_mesh, messages, stream
 from monarch.common.controller_api import TController
+from monarch.common.function import ResolvableFunction
 from monarch.common.invocation import Seq
 from monarch.common.messages import Referenceable, SendResultOfActorCall
 from monarch.common.stream import StreamRef
-from monarch.common.tensor import InputChecker, Tensor
+from monarch.common.tensor import dtensor_check, InputChecker, Tensor
+from monarch.common.tree import flatten
 from monarch.tensor_worker_main import _set_trace
 
 if TYPE_CHECKING:
@@ -265,17 +268,36 @@ class RemoteException(Exception):
             return "<exception formatting RemoteException>"
 
 
-def actor_send(
+def _cast_call_method_indirect(
     endpoint: ActorEndpoint,
+    selection: Selection,
+    client: MeshClient,
+    seq: Seq,
     args_kwargs_tuple: bytes,
     refs: Sequence[Any],
-    port: Optional[Port[Any]],
-    selection: Selection,
-):
+) -> Tuple[str, int]:
     unflatten_args = [
         UnflattenArg.PyObject if isinstance(ref, Tensor) else UnflattenArg.Mailbox
         for ref in refs
     ]
+    broker_id: Tuple[str, int] = client._mesh_controller.broker_id
+    actor_msg = PythonMessage(
+        PythonMessageKind.CallMethodIndirect(
+            endpoint._name, broker_id, seq, unflatten_args
+        ),
+        args_kwargs_tuple,
+    )
+    endpoint._actor_mesh.cast(actor_msg, selection)
+    return broker_id
+
+
+def actor_send(
+    endpoint: ActorEndpoint,
+    args_kwargs_tuple: bytes,
+    refs: Sequence[Any],
+    port: Optional[Port[Any]],
+    selection: Selection,
+):
     tensors = [ref for ref in refs if isinstance(ref, Tensor)]
     # we have some monarch references, we need to ensure their
     # proc_mesh matches that of the tensors we sent to it
@@ -284,7 +306,7 @@ def actor_send(
         if hasattr(t, "stream"):
             chosen_stream = t.stream
             break
-    with InputChecker(refs, lambda x: f"actor_call({x})") as checker:
+    with InputChecker(tensors, lambda x: f"actor_call({x})") as checker:
         checker.check_mesh_stream_local(device_mesh._active, chosen_stream)
         # TODO: move propagators into Endpoint abstraction and run the propagator to get the
         # mutates
@@ -300,8 +322,6 @@ def actor_send(
 
     client = cast(MeshClient, checker.mesh.client)
 
-    broker_id: Tuple[str, int] = client._mesh_controller.broker_id
-
     stream_ref = chosen_stream._to_ref(client)
 
     fut = (port, checker.mesh._ndslice) if port is not None else None
@@ -316,13 +336,9 @@ def actor_send(
     # The message to the generic actor tells it to first wait on the broker to get the local arguments
     # from the stream, then it will run the actor method, and send the result to response port.
 
-    actor_msg = PythonMessage(
-        PythonMessageKind.CallMethodIndirect(
-            endpoint._name, broker_id, ident, unflatten_args
-        ),
-        args_kwargs_tuple,
+    broker_id = _cast_call_method_indirect(
+        endpoint, selection, client, ident, args_kwargs_tuple, refs
     )
-    endpoint._actor_mesh.cast(actor_msg, selection)
     worker_msg = SendResultOfActorCall(ident, broker_id, tensors, [], stream_ref)
     client.send(checker.mesh._ndslice, worker_msg)
     # we have to ask for status updates
@@ -330,3 +346,49 @@ def actor_send(
     # enough work to count this future as finished,
     # and all potential errors have been reported
     client._request_status()
+
+
+def actor_rref(endpoint, args_kwargs_tuple: bytes, refs: Sequence[Any]):
+    chosen_stream = stream._active
+    fake_result, dtensors, mutates, mesh = dtensor_check(
+        endpoint._propagate,
+        cast(ResolvableFunction, endpoint._name),
+        refs,
+        {},
+        device_mesh._active,
+        chosen_stream,
+    )
+    assert mesh is not None
+
+    fake_result_dtensors, unflatten_result = flatten(
+        fake_result, lambda x: isinstance(x, torch.Tensor)
+    )
+    result_dtensors = tuple(
+        Tensor(fake, mesh, chosen_stream) for fake in fake_result_dtensors
+    )
+    seq = mesh.client.new_node(result_dtensors + mutates, dtensors)
+    assert all(t.ref is not None for t in result_dtensors)
+    assert all(t.ref is not None for t in mutates)
+    result = result_msg = unflatten_result(result_dtensors)
+    if len(result_dtensors) == 0:
+        result_msg = None
+
+    broker_id = _cast_call_method_indirect(
+        endpoint, "all", mesh.client, seq, args_kwargs_tuple, refs
+    )
+    # note the device mesh has to be defined regardles so the remote functions
+    # can invoke mesh.rank("...")
+
+    mesh.define_remotely()
+
+    mesh._send(
+        messages.CallActorMethod(
+            seq,
+            result_msg,
+            broker_id,
+            refs,
+            cast("List[Ref]", mutates),
+            stream._active._to_ref(mesh.client),
+        )
+    )
+    return result

monarch/tools/cli.py

@@ -86,9 +86,9 @@ class CreateCmd:
             else defaults.component_fn(config.scheduler)
         )
         component_args = component_args_from_cli(component_fn, args.component_args)
-        appdef = component_fn(**component_args)
+        config.appdef = component_fn(**component_args)
 
-        handle = create(config, appdef)
+        handle = create(config)
         print(handle)
 
 

monarch/tools/commands.py

@@ -7,18 +7,19 @@
 # pyre-strict
 
 import argparse
+import asyncio
 import inspect
 import logging
 import os
-import time
-from datetime import timedelta
+from datetime import datetime, timedelta
 from typing import Any, Callable, Mapping, Optional, Union
 
+from monarch.tools.components.hyperactor import DEFAULT_NAME
+
 from monarch.tools.config import (  # @manual=//monarch/python/monarch/tools/config/meta:defaults
     Config,
     defaults,
 )
-
 from monarch.tools.mesh_spec import mesh_spec_from_metadata, ServerSpec
 from torchx.runner import Runner  # @manual=//torchx/runner:lib_core
 from torchx.specs import AppDef, AppDryRunInfo, AppState, CfgVal, parse_app_handle
@@ -83,7 +84,7 @@ def component_args_from_cli(
 
 def create(
     config: Config,
-    appdef: AppDef,
+    name: str = DEFAULT_NAME,
 ) -> Union[str, AppDryRunInfo]:
     """Creates a monarch server by submitting it as a job to the target scheduler.
 
@@ -94,7 +95,7 @@ def create(
         from monarch.tools.config import defaults
 
         config = defaults.config(scheduler="slurm")
-        appdef = defaults.component_fn(scheduler=config.scheduler)()
+        config.appdef = defaults.component_fn(scheduler=config.scheduler)()
 
         config.scheduler_args.update(
             {
@@ -105,7 +106,7 @@ def create(
         )
         config.dryrun = True
 
-        create(config, appdef)
+        create(config)
 
 
     Args:
@@ -114,6 +115,7 @@ def create(
         component_fn: a function that returns the AppDef (job def).
             If not provided, defaults to the configured default for the scheduler
             (in most cases ``monarch.tools.components.hyperactor.proc_mesh``)
+        name: the name of the job. If none, a default job name will be created.
     """
     scheduler: str = config.scheduler
     cfg: Mapping[str, CfgVal] = config.scheduler_args
@@ -122,6 +124,8 @@ def create(
     os.environ["TORCHX_CONTEXT_NAME"] = os.getenv("TORCHX_CONTEXT_NAME", "monarch")
 
     with torchx_runner() as runner:
+        appdef: AppDef = AppDef(name, config.appdef.roles, config.appdef.metadata)
+
         info = runner.dryrun(appdef, scheduler, cfg, config.workspace)
 
         info_json_fmt = AppDryRunInfo(
@@ -170,6 +174,8 @@ def info(server_handle: str) -> Optional[ServerSpec]:
         # null-guard since some schedulers do not fill replica_status
         if host_status := replica_status.get(role.name):
             spec.hostnames = [h.hostname for h in host_status]
+            # the mesh status is based on the "least progressive" replica status
+            spec.state = min(h.state for h in host_status)
 
         mesh_specs.append(spec)
 
@@ -211,6 +217,8 @@ async def server_ready(
 
     """
 
+    check_interval_seconds = check_interval.total_seconds()
+    start = datetime.now()
     while True:
         server_spec = info(server_handle)
 
@@ -220,42 +228,56 @@ async def server_ready(
         if server_spec.state <= AppState.PENDING:  # UNSUBMITTED or SUBMITTED or PENDING
             # NOTE: TorchX currently does not have async APIs so need to loop-on-interval
             # TODO maybe inverse exponential backoff instead of constant interval?
-            check_interval_seconds = check_interval.total_seconds()
-            logger.info(
-                "waiting for %s to be %s (current: %s), will check again in %g seconds...",
-                server_handle,
-                AppState.RUNNING,
-                server_spec.state,
-                check_interval_seconds,
+            print(
+                f"Waiting for {server_handle} to be {AppState.RUNNING} (current: {server_spec.state}); "
+                f"will check again in {check_interval_seconds} seconds. "
+                f"Total wait time: {datetime.now() - start}",
+                end="\r",
             )
-            time.sleep(check_interval_seconds)
+            await asyncio.sleep(check_interval_seconds)
             continue
-        else:
-            return server_spec
-
 
+        # check if hosts are allocated for all the meshes
+        if server_spec.state == AppState.RUNNING:
+            running = True
+            for mesh_spec in server_spec.meshes:
+                if mesh_spec.state <= AppState.PENDING:
+                    print(
+                        f"Job {server_handle} is running but waiting for mesh {mesh_spec.name} "
+                        f"to be {AppState.RUNNING} (current: {mesh_spec.state}); "
+                        f"will check again in {check_interval_seconds} seconds. "
+                        f"Total wait time: {datetime.now() - start}",
+                        end="\r",
+                    )
+                    running = False
+                    break
+            if not running:
+                await asyncio.sleep(check_interval_seconds)
+                continue
+
+        return server_spec
+
+
+# TODO: this API is overloaded. Ideally, we do not need config to get or an handle to create.
 async def get_or_create(
     name: str,
     config: Config,
-    appdef: AppDef,
     check_interval: timedelta = _5_SECONDS,
 ) -> ServerSpec:
-    """Waits for the server called `name` in the scheduler specified in the `config`
+    """Waits for the server based on identity `name` in the scheduler specified in the `config`
     to be ready (e.g. RUNNING). If the server is not found then this function creates one
-    per the `appdef` spec, and waits for the server to be ready before returning.
+    per the `config` spec, and waits for the server to be ready before returning.
 
     Usage:
 
     .. code-block:: python
 
-        import getpass
         from monarch.tools.config import defaults
 
-        USER = getpass.getuser()
         config = defaults.config(scheduler)
-        appdef = defaults.component_fn(config.scheduler)()
+        config.appdef = defaults.component_fn(config.scheduler)()
 
-        server_handle = get_or_create(f"{USER}_monarch", config, appdef)
+        server_handle = get_or_create(name="my_job_name", config)
         server_info = info(server_handle)
 
     Returns: A `ServerSpec` containing information about either the existing or the newly
@@ -273,7 +295,7 @@ async def get_or_create(
         )
 
         # no dryrun (see assertion above) support so will always be a handle (str)
-        new_server_handle = str(create(config, appdef))
+        new_server_handle = str(create(config, name))
 
         logger.info(f"created new `{new_server_handle}` waiting for it to be ready...")
 
@@ -289,10 +311,10 @@ async def get_or_create(
                 f"the new server `{new_server_handle}` has {server_info.state}"
             )
 
-        logger.info(f"server `{new_server_handle}` is: {server_info.state}")
+        print(f"\x1b[36mNew job `{new_server_handle}` is ready to serve. \x1b[0m")
         return server_info
     else:
-        logger.info("found existing RUNNING server `%s`", server_handle)
+        print(f"\x1b[36mFound existing job `{server_handle}` ready to serve. \x1b[0m")
         return server_info
 
 

monarch/tools/components/hyperactor.py

@@ -9,6 +9,7 @@ import getpass
 from typing import Optional
 
 from monarch.tools import mesh_spec
+from monarch.tools.config import UnnamedAppDef
 from monarch.tools.mesh_spec import mesh_spec_from_str
 from torchx import specs
 
@@ -16,17 +17,18 @@ _DEFAULT_MESHES = ["mesh_0:1:gpu.small"]
 
 _USER: str = getpass.getuser()
 
+DEFAULT_NAME: str = f"monarch-{_USER}"
+
 __version__ = "latest"  # TODO get version from monarch.__version_
 
 
 def proc_mesh(
-    name: str = f"monarch-{_USER}",
     image: str = f"ghcr.io/pytorch-labs/monarch:{__version__}",  # TODO docker needs to be built and pushed to ghcr
     meshes: list[str] = _DEFAULT_MESHES,
     env: Optional[dict[str, str]] = None,
     port: int = mesh_spec.DEFAULT_REMOTE_ALLOCATOR_PORT,
     program: str = "monarch_bootstrap",  # installed with monarch wheel (as console script)
-) -> specs.AppDef:
+) -> UnnamedAppDef:
     """
     Args:
         name: the name of the monarch server job
@@ -37,7 +39,7 @@ def proc_mesh(
         program: path to the binary that the remote process allocator spawns on an allocation request
     """
 
-    appdef = specs.AppDef(name)
+    appdef = UnnamedAppDef()
 
     for mesh in [mesh_spec_from_str(mesh) for mesh in meshes]:
         mesh_role = specs.Role(

monarch/tools/config/__init__.py

@@ -6,15 +6,32 @@
 
 # pyre-strict
 from dataclasses import dataclass, field
-from typing import Any, Optional
+from typing import Any, Dict, List, Optional
+
+from torchx.specs import Role
 
 
 NOT_SET: str = "__NOT_SET__"
 
 
+@dataclass
+class UnnamedAppDef:
+    """
+    A TorchX AppDef without a name.
+    """
+
+    roles: List[Role] = field(default_factory=list)
+    metadata: Dict[str, str] = field(default_factory=dict)
+
+
 @dataclass
 class Config:
+    """
+    All configs needed to schedule a mesh of allocators.
+    """
+
     scheduler: str = NOT_SET
     scheduler_args: dict[str, Any] = field(default_factory=dict)
     workspace: Optional[str] = None
     dryrun: bool = False
+    appdef: UnnamedAppDef = UnnamedAppDef()

monarch/tools/config/defaults.py

@@ -11,7 +11,7 @@
 from typing import Callable, Optional
 
 from monarch.tools.components import hyperactor
-from monarch.tools.config import Config
+from monarch.tools.config import Config, UnnamedAppDef
 
 from torchx import specs
 from torchx.schedulers import (
@@ -23,7 +23,7 @@ from torchx.schedulers import (
 )
 
 
-def component_fn(scheduler: str) -> Callable[..., specs.AppDef]:
+def component_fn(scheduler: str) -> Callable[..., UnnamedAppDef]:
     """The default TorchX component function for the scheduler"""
     return hyperactor.proc_mesh
 

monarch/tools/mesh_spec.py

@@ -9,6 +9,8 @@ import string
 from dataclasses import dataclass, field
 from typing import Any, Optional
 
+from monarch.tools.config import UnnamedAppDef
+
 from monarch.tools.network import get_sockaddr
 from torchx import specs
 from torchx.specs.api import is_terminal
@@ -39,6 +41,7 @@ class MeshSpec:
     transport: str = "tcp"
     port: int = DEFAULT_REMOTE_ALLOCATOR_PORT
     hostnames: list[str] = field(default_factory=list)
+    state: specs.AppState = specs.AppState.UNSUBMITTED
 
     def server_addrs(
         self, transport: Optional[str] = None, port: Optional[int] = None
@@ -69,7 +72,7 @@ def _tag(mesh_name: str, tag_template: str) -> str:
     return string.Template(tag_template).substitute(mesh_name=mesh_name)
 
 
-def tag_as_metadata(mesh_spec: MeshSpec, appdef: specs.AppDef) -> None:
+def tag_as_metadata(mesh_spec: MeshSpec, appdef: UnnamedAppDef) -> None:
     appdef.metadata[_tag(mesh_spec.name, _TAG_HOST_TYPE)] = mesh_spec.host_type
     appdef.metadata[_tag(mesh_spec.name, _TAG_GPUS)] = str(mesh_spec.gpus)
     appdef.metadata[_tag(mesh_spec.name, _TAG_TRANSPORT)] = mesh_spec.transport

tests/test_allocator.py

@@ -33,7 +33,6 @@ from monarch._rust_bindings.monarch_hyperactor.channel import (
     ChannelTransport,
 )
 
-from monarch._src.actor.actor_mesh import MonarchContext
 from monarch._src.actor.allocator import (
     ALLOC_LABEL_PROC_MESH_NAME,
     LocalAllocator,
@@ -160,7 +159,7 @@ class TestSetupActorInAllocator(unittest.IsolatedAsyncioTestCase):
             "TEST_ENV_VAR_3": "value_3",
         }
 
-        def setup_multiple_env_vars(ctx: MonarchContext) -> None:
+        def setup_multiple_env_vars() -> None:
             for name, value in env_vars.items():
                 os.environ[name] = value
 
@@ -184,36 +183,33 @@ class TestSetupActorInAllocator(unittest.IsolatedAsyncioTestCase):
             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"
+        """Test that the setup lambda can access rank information"""
+        context_var_name: str = "PROC_MESH_RANK_INFO"
 
-        def setup_with_context(ctx: MonarchContext) -> None:
-            context_info = f"proc_id:{ctx.proc_id},point_rank:{ctx.point.rank}"
+        def setup_with_rank() -> None:
+            context_info = f"point_rank:{current_rank().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)
+        proc_mesh = await ProcMesh.from_alloc(alloc, setup=setup_with_rank)
 
         try:
             actor = await proc_mesh.spawn("env_check", EnvCheckActor)
 
-            context_info = await actor.get_env_var.call_one(context_var_name)
+            rank_info = await actor.get_env_var.call_one(context_var_name)
 
             self.assertNotEqual(
-                context_info,
+                rank_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",
+                rank_info,
+                f"Context information {rank_info} does not contain point_rank",
             )
         finally:
             await proc_mesh.stop()
@@ -435,7 +431,7 @@ class TestRemoteAllocator(unittest.IsolatedAsyncioTestCase):
         test_var_name: str = "TEST_ENV_VAR_FOR_PROC_MESH"
         test_var_value: str = "test_value_123"
 
-        def setup_env_vars(ctx: MonarchContext) -> None:
+        def setup_env_vars() -> None:
             os.environ[test_var_name] = test_var_value
 
         hosts = 2

tests/test_env_before_cuda.py

@@ -15,7 +15,6 @@ import cloudpickle
 
 import torch
 from monarch._rust_bindings.monarch_hyperactor.alloc import AllocConstraints, AllocSpec
-from monarch._src.actor.actor_mesh import MonarchContext
 from monarch._src.actor.allocator import LocalAllocator
 from monarch._src.actor.proc_mesh import proc_mesh
 from monarch.actor import Actor, endpoint, ProcMesh
@@ -70,7 +69,7 @@ class TestEnvBeforeCuda(unittest.IsolatedAsyncioTestCase):
             "CUDA_LAUNCH_BLOCKING": "1",
         }
 
-        def setup_cuda_env(_: MonarchContext) -> None:
+        def setup_cuda_env() -> None:
             for name, value in cuda_env_vars.items():
                 os.environ[name] = value
 
@@ -107,7 +106,7 @@ class TestEnvBeforeCuda(unittest.IsolatedAsyncioTestCase):
             "CUDA_DEVICE_MAX_CONNECTIONS": "1",
         }
 
-        def setup_cuda_env(_: MonarchContext) -> None:
+        def setup_cuda_env() -> None:
             for name, value in cuda_env_vars.items():
                 os.environ[name] = value
 

tests/test_tensor_engine.py

@@ -8,7 +8,7 @@ import monarch
 import pytest
 import torch
 from monarch import remote
-from monarch.actor import Actor, endpoint, proc_mesh
+from monarch.actor import Actor, as_endpoint, endpoint, proc_mesh
 from monarch.mesh_controller import spawn_tensor_engine
 
 
@@ -104,3 +104,29 @@ def test_actor_tensor_ordering() -> None:
             results.append(counter.incr.call(1))
 
         assert list(range(10)) == [r.get().item(hosts=0, gpus=0) for r in results]
+
+
+class Linear(Actor):
+    def __init__(self, N: int, M: int):
+        self.weight = torch.zeros((N, M))
+
+    def forward(self, x) -> torch.Tensor:
+        return x @ self.weight
+
+    @endpoint(propagate="inspect")
+    def update(self, w: torch.Tensor) -> None:
+        self.weight += w
+
+
+@two_gpu
+def test_rref_actor() -> None:
+    pm = proc_mesh(gpus=1).get()
+    with pm.activate():
+        x = pm.spawn("linear", Linear, 3, 4).get()
+
+        y = torch.ones((4, 3))
+        t = as_endpoint(x.forward, propagate=lambda x: torch.rand(3, 4)).rref(y)
+        assert monarch.inspect(t.sum()).item() == 0
+        x.update.rref(torch.ones((3, 4)))
+        t = as_endpoint(x.forward, propagate=lambda x: torch.rand(3, 4)).rref(y)
+        assert monarch.inspect(t.sum()).item() == 3 * 4 * 4

{torchmonarch_nightly-2025.7.25.dist-info → torchmonarch_nightly-2025.7.26.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: torchmonarch-nightly
-Version: 2025.7.25
+Version: 2025.7.26
 Summary: Monarch: Single controller library
 Author: Meta
 Author-email: oncall+monarch@xmail.facebook.com
@@ -44,6 +44,8 @@ Note: Monarch is currently only supported on Linux systems
 
 ## Installation
 
+### On Fedora distributions
+
 `pip install torchmonarch-nightly`
 
 or manually
@@ -88,6 +90,37 @@ pip install --no-build-isolation -e .
 pytest python/tests/ -v -m "not oss_skip"
 ```
 
+### On MacOS
+
+You can also build Monarch to run locally on a MacOS system.
+
+Note that this does not support tensor engine, which is tied to CUDA and RDMA (via ibverbs).
+
+
+```sh
+
+# Create and activate the conda environment
+conda create -n monarchenv python=3.10 -y
+conda activate monarchenv
+
+# Install nightly rust toolchain
+curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
+rustup toolchain install nightly
+rustup default nightly
+
+# Install build dependencies
+pip install -r build-requirements.txt
+# Install test dependencies
+pip install -r python/tests/requirements.txt
+
+# Build and install Monarch
+USE_TENSOR_ENGINE=0 pip install --no-build-isolation .
+# or setup for development
+USE_TENSOR_ENGINE=0 pip install --no-build-isolation -e .
+
+```
+
+
 ## Running examples
 
 Check out the `examples/` directory for demonstrations of how to use Monarch's APIs.

{torchmonarch_nightly-2025.7.25.dist-info → torchmonarch_nightly-2025.7.26.dist-info}/RECORD

@@ -1,5 +1,5 @@
 monarch/__init__.py,sha256=mgKiyD1kxky-1pvhMlNfF4VmxWnhi-FSYZNFzkW1BEM,7052
-monarch/_rust_bindings.so,sha256=nHFmFZoCgxS1M40RBTn52oOVV9nR7eJ88hwz0jf_g6c,47433680
+monarch/_rust_bindings.so,sha256=pvUGrYVAW1JrVgti7IJsFSb7Va0vWTB95RvoOgAIXGE,47520848
 monarch/_testing.py,sha256=_3MYNMq-_0T1qXCj2vxrW13GlWGdUuVFMskQF2Gsw_o,7877
 monarch/actor_mesh.py,sha256=VtPU9syi_vUdwDSJJ639Z4Y_EcWZUScyoj0lQ88RQPs,421
 monarch/bootstrap_main.py,sha256=39OZpNMrfvvNJf-iwuNzgslzYA_ItaRPHfXGn_V74N0,524
@@ -7,8 +7,8 @@ monarch/cached_remote_function.py,sha256=kYdB6r4OHx_T_uX4q3tCNcp1t2DJwF8tPTIahUi
 monarch/fetch.py,sha256=JMxC8HdWMvpik0T4E6e-gfxvmNnOkA0ul4eo4R3Jg_o,1712
 monarch/gradient_generator.py,sha256=Rl3dmXGceTdCc1mYBg2JciR88ywGPnW7TVkL86KwqEA,6366
 monarch/memory.py,sha256=ol86dBhFAJqg78iF25-BuK0wuwj1onR8FIioZ_B0gjw,1377
-monarch/mesh_controller.py,sha256=uFAExVNzMGoeQI0kmuv4-kMbKHn420oeaG1rTYFi2sg,11884
-monarch/monarch_controller,sha256=bHrKU_rB52VnriwATjytMs8vmeLmZfFC04B7g0X9aLE,21202456
+monarch/mesh_controller.py,sha256=mOk2misobJun2AgR_ALjFoopAEcOPYQVrrAJXa18ZTs,13810
+monarch/monarch_controller,sha256=1gXdTNolxIUp_88alJnW1onBHseo5PzuNnjEnBte4wI,21243128
 monarch/notebook.py,sha256=zu9MKDFKf1-rCM2TqFSRJjMBeiWuKcJSyUFLvoZRQzs,25949
 monarch/opaque_module.py,sha256=jCcg0DjbcEVXA9WNG0NhUzGteLHOJLTZEBvrIYJIAns,10436
 monarch/opaque_object.py,sha256=x1LoX6RIMGh4ux52xIfhPgoh6PhZHdkf9bMccHW3DW0,2808
@@ -25,27 +25,27 @@ monarch/tensorboard.py,sha256=MnLgH5lbqeUJauEuirEgR6L_qYl2NGdtwZOWIAuOZao,2587
 monarch/world_mesh.py,sha256=ob5dJWaC49Uw0xqClHBm8CQLvL4xKnjd4TGzk7k8NxI,980
 monarch/_src/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 monarch/_src/actor/__init__.py,sha256=4iK3nzQZmEPe0HGNY70fABBenw3lCVVaaF0xddF5Fa0,235
-monarch/_src/actor/actor_mesh.py,sha256=nG56lE6RwcNHeF4SUutf1NPmO8GQMaIjCjphCxf_pRU,29233
+monarch/_src/actor/actor_mesh.py,sha256=guYD9nZHguLGJAvTisc3Q664ASkupcNC6z9iheeGFUQ,29188
 monarch/_src/actor/allocator.py,sha256=WpHEK1SvjgF3GdIWIYUkonXli2-gQVKJVZPInl2RFQo,8212
 monarch/_src/actor/bootstrap_main.py,sha256=e5eU3GvX60MWWmCty7VcZrAmukD29iJdWBysNgQ2o3A,2342
 monarch/_src/actor/debugger.py,sha256=t2iAAxz03b2KZ89T3VjRc__7GzSf83R8gM81SDyX3-o,19532
 monarch/_src/actor/device_utils.py,sha256=gBpl23wMjppVAEzzj8U9HyX-B7Bs2_3ftiMAkzUS4j4,577
-monarch/_src/actor/endpoint.py,sha256=V8J4LVTnTFeo4HhkdhISmepzJOxdqbHe6cxwYkj9Qyg,8462
-monarch/_src/actor/event_loop.py,sha256=gpfbPoOhrY8W2Z5cDcujIfGI0NTdLrRycs7FmbsVatE,2830
+monarch/_src/actor/endpoint.py,sha256=jM3XYWze6gB6ajE4AMojNFSN4IaaxgioNAErJkkywjE,9721
+monarch/_src/actor/event_loop.py,sha256=2i4fKIkemBzua_t47BqVa2roZ6fWB6sbmMFPNx2zKN0,2832
 monarch/_src/actor/future.py,sha256=jOGh1wfwKyGJxhl9t1P8eapXYywf8KwQldZCCbupmb8,4042
 monarch/_src/actor/pdb_wrapper.py,sha256=-QxRktntdEO2LdHixBGKLboYtADyh8bEIAoa3gFwIEo,4161
 monarch/_src/actor/pickle.py,sha256=jD_3E07OJmMIlcMOOrNFnIuRKZU2F_Q_BP-njDFXUNM,2044
-monarch/_src/actor/proc_mesh.py,sha256=mbXgoUAnajKqe54LQSV3QfWii2b28-NxL4YmbpS4hxM,16368
+monarch/_src/actor/proc_mesh.py,sha256=amF4fbO-33qHFudlS9WabYXIVh0Y_D_0nhCTxvOhpGg,16640
 monarch/_src/actor/shape.py,sha256=B-7DI768ZhT8ECUNCJcI7DfCB7iDFGFH0r-HmXaAfcM,8296
 monarch/_src/actor/sync_state.py,sha256=GB6bTAGztkcN8fZ9K7zXklOzjYzv6cvkJeBje20xFkE,471
-monarch/_src/actor/tensor_engine_shim.py,sha256=r9ZK0ELVvn-cbs4LdP7IxO9KZaLV9p6q36ERbthjEd0,1568
+monarch/_src/actor/tensor_engine_shim.py,sha256=hupavQ2rjPwECaTlDAhY-eeiEY18Wyyx59MZHcSEcYM,1622
 monarch/_src/actor/code_sync/__init__.py,sha256=qzWoFNJEJvEbqab0QuHbkvhdz6FHi7BOTw6-2B3p0A4,378
 monarch/_src/actor/code_sync/auto_reload.py,sha256=kqXCQuSzjxMw8bcDLsUZiL_NImo4j2EScfNklwpltmU,6685
 monarch/_src/actor/telemetry/__init__.py,sha256=sHA5fmFdWU9jcUJVszNFhbXbjRSIBmuDXDMwJrrE0hw,523
 monarch/_src/actor/telemetry/rust_span_tracing.py,sha256=UvkywuwjQX7tIyLdKZbF-fcmI_aHporAejsTRTyJNNg,4445
 monarch/_src/tensor_engine/__init__.py,sha256=Md3cCHD7Ano9kV15PqGbicgUO-RMdh4aVy1yKiDt_xE,208
 monarch/_src/tensor_engine/rdma.py,sha256=KbhJXMuuHruYXnmxzB3BpkpcGsZ4hSu_7C6wF-EPhDk,6331
-monarch/actor/__init__.py,sha256=oQY131pVHg9iBnyZIi-zHvDturO6VuSlfnWktLwI3cE,1014
+monarch/actor/__init__.py,sha256=F87BC7owDdH_yRjLvMu6pbICbajndsEbtWG2W53Rapo,1050
 monarch/builtins/__init__.py,sha256=QcfnHZGbc2qktBg7DyZt2ruE6VahnIt4S8lEZLHdJqU,443
 monarch/builtins/log.py,sha256=H1QkuVzwxyi36Zyv-XR0VN0QsNimBWwxE1__fjs0_2o,554
 monarch/builtins/random.py,sha256=wPbvscg7u53EXpMFo885fO2XOlsyjrNAJ4rBxLzfxdg,1839
@@ -67,14 +67,14 @@ monarch/common/function_caching.py,sha256=HVdbWtv6Eea7ENMWi8iv36w1G1TaVuUJhkUX_J
 monarch/common/future.py,sha256=D1UJ_8Rvb8-VG9vNE-z7xz2m2otMd2HgB0rnA02nlvA,4681
 monarch/common/invocation.py,sha256=L4mSmzqlHMxo1Tb71hBU_M8aBZCRCOcb6vvPhvvewec,4195
 monarch/common/mast.py,sha256=XTzYljGR0aZ7GjmNMPgU2HyuL4HWSAy4IwE3kEDqdOw,7735
-monarch/common/messages.py,sha256=Xp2TdVhCN52qpSnDGzi1o813okYZ9-vY9mBxw7ZFGVg,18606
+monarch/common/messages.py,sha256=jwwJMVVx3gKd39AXcnRxjMr7lPJRLimHtZYel3zjq4o,18833
 monarch/common/mock_cuda.py,sha256=x6ho1Ton6BbKjBZ5ZxnFOUaQM032X70wnpoUNB7Ci2w,1039
 monarch/common/opaque_ref.py,sha256=tWNvOC6CsjNPKD1JDx-8PSaeXqZC3eermgBExUPKML4,2871
 monarch/common/pipe.py,sha256=9pTf8--3yOv4HpnJEhgcmc_JM6Az4uL1y72TSQA55dw,5013
 monarch/common/process_group.py,sha256=FbJ_AJRZYFkvQ68L2naRq64J_aNuAKe5kO0MWdn_x74,1662
 monarch/common/recording.py,sha256=Q39Zhb3kT52NCPf4VVMox2WXjtXju5eTuvPMZ_QGW7o,4660
 monarch/common/reference.py,sha256=O26lkzEeVwj0S1xEy-OLqdHVnACmmlbQCUmXRrW4n1Q,938
-monarch/common/remote.py,sha256=h0niT9wDHuRuIPwQYBa4QVM803BxA91ANjsnKH2ZyXc,12144
+monarch/common/remote.py,sha256=Q2YpAo_fsdh22ElUNruxYyn-cNTecZr8POcHCGtuKyg,12129
 monarch/common/selection.py,sha256=lpWFbZs3ArYy29e-53eoAVAjQFksf1RvZz9NvM0CUW4,308
 monarch/common/stream.py,sha256=_ejoxafHtdD10lLzznRCXKwrkZ_ZH9k_VTgiA5yfBrI,3583
 monarch/common/tensor.py,sha256=ysT51NClNF4FxV0DFLJJUNmCRaVy8uQuYWpLViyPLdY,29292
@@ -89,7 +89,7 @@ monarch/controller/rust_backend/__init__.py,sha256=J8qjUOysmcMAek2KFN13mViOXZxTY
 monarch/controller/rust_backend/controller.py,sha256=8IYnVUiqEVKO9rGL3vKqcCSAhWJG1bYYQ0MoaMqsp78,9521
 monarch/gradient/__init__.py,sha256=kqmzwt16mMpk0M3GhpgP_f7da4DGnaV9chDzbt66k4Q,308
 monarch/gradient/_gradient_generator.pyi,sha256=6cX0UxaDt9NAlwgIhTgnweqGOf6qRhHiGnUzSWNCxdU,630
-monarch/gradient/_gradient_generator.so,sha256=WyGDzhfswBpL-pzHSG_FQVgOS-ZMTTk34gswjFtt8ZU,11531728
+monarch/gradient/_gradient_generator.so,sha256=VyBW9SsE1IJ5iiiq1Su0BgTR7vDBnbaRQj7yjAtBPUs,11531728
 monarch/parallel/__init__.py,sha256=6920kIkhiX7AiyjYvyc1ad8ccP-bStJJ1sS5KkeN2P0,352
 monarch/parallel/pipelining/__init__.py,sha256=J8qjUOysmcMAek2KFN13mViOXZxTYc5vCrF02t3VuFU,223
 monarch/parallel/pipelining/runtime.py,sha256=KK8TG1gUYEzSsquiZoPTWGSIC74mlncD7cYknKxfb3c,32470
@@ -115,15 +115,15 @@ monarch/timer/example_spmd.py,sha256=p8i3_tO1AmpwSkZryiSjgkh7qaEZ6QXp2Fy1qtPpECA
 monarch/timer/execution_timer.py,sha256=1YsrLIZirdohKOeFAU2H4UcONhQXHuctJbYcoX8I6gY,6985
 monarch/timer/execution_timer_test.py,sha256=CSxTv44fFZQURJlCBmYvysQI1aS_zEGZs_uxl9SOHak,4486
 monarch/tools/__init__.py,sha256=J8qjUOysmcMAek2KFN13mViOXZxTYc5vCrF02t3VuFU,223
-monarch/tools/cli.py,sha256=rni8Et8_uMZLTKwwdqeFTia84pDwMh6Ne-IJx36LYsc,5002
-monarch/tools/commands.py,sha256=fU4EPnNx0M2iH4eLJsMqPDzIGl6t9ALSJc1szLHQP6Y,10821
-monarch/tools/mesh_spec.py,sha256=kAo_GcU0LOm6cBkbeBGBtU9WKKS0kiDG-M7Uf9Wrp7Y,7831
+monarch/tools/cli.py,sha256=b3mKZnK-MwP7JwskTxHI0KcJXxSU6498jEb2ntVr_VM,5001
+monarch/tools/commands.py,sha256=3xuvHcMwl0t6cWTVUxI_r8EqrJZnay0bkKxOijhlKrw,12126
+monarch/tools/mesh_spec.py,sha256=in6txNRmA-UvveVSMHCjX6mGpofd3K8vl2Plz1eD6rg,7935
 monarch/tools/network.py,sha256=mN8Fx9mervxM3VdFHRn4ZXt4z7yWxZp52BTxx2tfpus,2455
 monarch/tools/utils.py,sha256=2GGUQQE0dLtzoKy40_tAsOfbSxE6krnL0WvwMgUBgmw,1213
 monarch/tools/components/__init__.py,sha256=J8qjUOysmcMAek2KFN13mViOXZxTYc5vCrF02t3VuFU,223
-monarch/tools/components/hyperactor.py,sha256=Ryi1X07VLcaQVlpc4af65JNBbZtOb9IAlKxSKMZ1AW4,2120
-monarch/tools/config/__init__.py,sha256=OPSflEmJB2zxAaRVzzWSWXV5M5vlknLgpulGdW1ze5U,510
-monarch/tools/config/defaults.py,sha256=34a3HQhyXqt9qR2SYMVCROoNsnwk37rIwLXXiKwqtog,1894
+monarch/tools/components/hyperactor.py,sha256=gYZS8AcmoTuq48mRrZWWnyxQqaiwTNHv8YqHhHi799U,2169
+monarch/tools/config/__init__.py,sha256=MLa6uvVJssN_zTciCvCMeCURWglchCuqE3zdqA-gh4U,869
+monarch/tools/config/defaults.py,sha256=ZymyKr9fNnBIgsV-xz-KrtrXRLkJo3hymTqxjXXnBzs,1910
 monarch/worker/__init__.py,sha256=J8qjUOysmcMAek2KFN13mViOXZxTYc5vCrF02t3VuFU,223
 monarch/worker/_testing_function.py,sha256=A81cVMKgdlO66XvoYcBCDrxIQIm3o3GgvcH_c8M9OmI,13480
 monarch/worker/compiled_block.py,sha256=hYx1F6PAu0_BnpKAprP_nV9qJtk5XWO7mcwH3JPDioU,10114
@@ -146,12 +146,12 @@ tests/error_test_binary.py,sha256=cgdrnVI3SIzAFSRXTvASfiR8eKSMrZ7N3tSCLVkJo44,78
 tests/sleep_binary.py,sha256=XfLYaAfwm9xgzM-svs8fhAeFhwYIg6SyVEnx4e6wbUw,1009
 tests/test_actor_error.py,sha256=kEfj1XW_WPk2mplucFBuzCWU3UrvzwkKoHSLIZfyQr0,20945
 tests/test_alloc.py,sha256=IW7yJSaKxhOYc8SJtFyREakDUwiKWq9M0CGgYyBYHoc,743
-tests/test_allocator.py,sha256=gETCLy7kMNVo17gxR3rnMq8kdH5IwcF2tVher-fAOxU,29047
+tests/test_allocator.py,sha256=4LcUB4QRNGDp0qBWAyLM6ektmoxpO922f-NcHZziJ_w,28762
 tests/test_coalescing.py,sha256=JZ4YgQNlWWs7N-Z8KCCXQPANcuyyXEKjeHIXYbPnQhk,15606
 tests/test_controller.py,sha256=CIMb-ApmBcBj1eCqccDUAbVyyJWMGooAha5gQk0AoeY,31452
 tests/test_debugger.py,sha256=mtd_no7dAooBePoQ_TZOxtgzwd1-x6xkpzAFK1_Y8B8,13703
 tests/test_device_mesh.py,sha256=DrbezYOM0thfP9MgLXb5-F0VoLOmSz5GR0GwjR_3bE4,5290
-tests/test_env_before_cuda.py,sha256=w00oi9aP0tFuZtUFggzA9h6qWXXgLo1rN1GoLJZbG10,5531
+tests/test_env_before_cuda.py,sha256=K5zdpXNRZB8hXQJaTN_CftcGHb3vzzdKasu8KFUoiCg,5440
 tests/test_fault_tolerance.py,sha256=u4wmG1z5MZ6PY6us5zUZHJh2pUC3L7i0wsUfRDNHmxA,14144
 tests/test_future.py,sha256=cXzaNi2YDwVyjR541ScXmgktX1YFsKzbl8wep0DMVbk,3032
 tests/test_grad_generator.py,sha256=p4Pm4kMEeGldt2jUVAkGKCB0mLccKI28pltH6OTGbQA,3412
@@ -163,15 +163,15 @@ tests/test_remote_functions.py,sha256=VT65W7htp1jCsP9-AsiO1dofhx-4OebWEOssDEgi3G
 tests/test_rust_backend.py,sha256=3TLu8dSVEqyLhjHED2DoAEQHTpbBQcr3WI6K2eGZLZw,7861
 tests/test_signal_safe_block_on.py,sha256=bmal0XgzJowZXJV6T1Blow5a-vZluYWusCThLMGxyTE,3336
 tests/test_sim_backend.py,sha256=kT7MnPo5O9xxX8f7uZOpR9Tkuz5brjaOyK1g1NqHRlo,1398
-tests/test_tensor_engine.py,sha256=LIJOb6hPVCpgLJjjPlcH2MgLIyM1JG7d-qMFpIUvFuQ,2793
+tests/test_tensor_engine.py,sha256=_F70SQiUCRVZcbq5JcP5XkGJFnul57pqBpu1rF9kipE,3591
 tests/simulator/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 tests/simulator/test_profiling.py,sha256=TGYCfzTLdkpIwnOuO6KApprmrgPIRQe60KRX3wkB0sg,4565
 tests/simulator/test_simulator.py,sha256=LO8lA0ssY-OGEBL5ipEu74f97Y765TEwfUOv-DtIptM,14568
 tests/simulator/test_task.py,sha256=ipqBDuDAysuo1xOB9S5psaFvwe6VATD43IovCTSs0t4,2327
 tests/simulator/test_worker.py,sha256=QrWWIJ3HDgDLkBPRc2mwYPlOQoXQcj1qRfc0WUfKkFY,3507
-torchmonarch_nightly-2025.7.25.dist-info/licenses/LICENSE,sha256=e0Eotbf_rHOYPuEUlppIbvwy4SN98CZnl_hqwvbDA4Q,1530
-torchmonarch_nightly-2025.7.25.dist-info/METADATA,sha256=sumslJxhBqMCVPa_3AoGQh7fY7dJHR6DhInBfTc6skE,3063
-torchmonarch_nightly-2025.7.25.dist-info/WHEEL,sha256=JC9FVdjbTDi9l3EyrqUd11CgmN9LkBi1g5dFHayafwA,104
-torchmonarch_nightly-2025.7.25.dist-info/entry_points.txt,sha256=60QVSpYVzkzS4iDOiLp0fsLxVp47X3J2l3v7W-59LMo,117
-torchmonarch_nightly-2025.7.25.dist-info/top_level.txt,sha256=E-ZssZzyM17glpVrh-S9--qJ-w9p2EjuYOuNw9tQ4Eg,33
-torchmonarch_nightly-2025.7.25.dist-info/RECORD,,
+torchmonarch_nightly-2025.7.26.dist-info/licenses/LICENSE,sha256=e0Eotbf_rHOYPuEUlppIbvwy4SN98CZnl_hqwvbDA4Q,1530
+torchmonarch_nightly-2025.7.26.dist-info/METADATA,sha256=rJuOVCi7kVf2R9tHhtMSlaWv80ybWS5g9MvMzmso5M8,3852
+torchmonarch_nightly-2025.7.26.dist-info/WHEEL,sha256=JC9FVdjbTDi9l3EyrqUd11CgmN9LkBi1g5dFHayafwA,104
+torchmonarch_nightly-2025.7.26.dist-info/entry_points.txt,sha256=60QVSpYVzkzS4iDOiLp0fsLxVp47X3J2l3v7W-59LMo,117
+torchmonarch_nightly-2025.7.26.dist-info/top_level.txt,sha256=E-ZssZzyM17glpVrh-S9--qJ-w9p2EjuYOuNw9tQ4Eg,33
+torchmonarch_nightly-2025.7.26.dist-info/RECORD,,