documentation

Author: allenwang28

apps/grpo/main.py

@@ -363,6 +363,9 @@ async def main(cfg: DictConfig):
 
     # In the HostMesh v1 case, we spawn a torchstore storage volume
     # per trainer process.
+    # We initialize after service initialization because torchstore currently
+    # requires access to the underlying proc meshes in the local rank strategy.
+    # We should be able to hide this in the future.
     if MONARCH_HOSTMESH_V1.get_value():
         # TODO: support multiple host meshes
         trainer_num_procs = cfg.actors.trainer["procs"]

src/forge/controller/launcher.py

@@ -121,6 +121,9 @@ async def remote_setup(self, procs: ProcMesh) -> tuple[str, int]:
 class Slurmlauncher(BaseLauncher):
     async def initialize(self) -> None:
         if MONARCH_HOSTMESH_V1.get_value():
+            # HostMeshV1 currently requires explicit configuration
+            # of the underlying transport from client to mesh.
+            # This can be removed in the future once this has been removed.
             configure(default_transport=ChannelTransport.Tcp)
 
     async def get_allocator(self, name: str, num_hosts: int) -> tuple[Any, Any, str]:
@@ -178,6 +181,9 @@ def __init__(self, cfg: LauncherConfig | None = None):
 
     async def initialize(self) -> None:
         if MONARCH_HOSTMESH_V1.get_value():
+            # HostMeshV1 currently requires explicit configuration
+            # of the underlying transport from client to mesh.
+            # This can be removed in the future once this has been removed.
             configure(default_transport=ChannelTransport.MetaTlsWithHostname)
 
         await self.launch_mast_job()

src/forge/controller/provisioner.py

@@ -183,6 +183,12 @@ async def create_host_mesh(self, name: str, num_hosts: int) -> HostMesh:
         return host_mesh, server_name
 
     def get_host_mesh(self, name: str) -> HostMesh:
+        """Returns the host mesh given its associated name.
+
+        This is currently an experimental API for HostMesh v1 and
+        should not be relied on longer term.
+
+        """
         return self._host_mesh_map[name]
 
     async def get_proc_mesh(