more updates for testing

Author: allenwang28

assets/wheels/monarch-0.0.1-cp310-cp310-linux_x86_64.whl

@@ -17,7 +17,7 @@ NC='\033[0m'
 # Configuration
 PYTORCH_VERSION="2.9.0.dev20250905"
 VLLM_BRANCH="v0.10.0"
-MONARCH_COMMIT="6ca383aca99480aa1bf5853478d4d09fcb224035"
+MONARCH_COMMIT="d1c5ea4732704454efad82db678d4e66a4131bb2"
 TORCHTITAN_COMMIT="0cfbd0b3c2d827af629a107a77a9e47229c31663"
 TORCHSTORE_COMMIT="eed96eb55ce87d4a9880597dd7dfd0d291e9ac81"
 BUILD_DIR="$HOME/forge-build"

scripts/build_wheels.sh

@@ -53,6 +53,7 @@
 from forge.data.sharding import VLLMSharding
 from forge.data_models.completion import Completion
 from forge.data_models.prompt import to_prompt
+from forge.env import TORCHSTORE_USE_RDMA
 from forge.interfaces import Policy as PolicyInterface
 from forge.observability.metrics import record_metric, Reduce
 from forge.observability.perf_tracker import Tracer
@@ -140,7 +141,7 @@ class Policy(PolicyInterface):
     engine_config: EngineConfig | Mapping = field(default_factory=EngineConfig)
     sampling_config: SamplingConfig | Mapping = field(default_factory=SamplingConfig)
     available_devices: str | None = None
-    use_dcp: bool = True
+    use_dcp: bool = not TORCHSTORE_USE_RDMA.get_value()
     # Gets set up by setup
     sampling_params: SamplingParams | None = None
     lora_request: LoRARequest | None = None

src/forge/actors/policy.py

@@ -46,6 +46,7 @@
 
 from forge.controller import ForgeActor
 from forge.data.utils import batch_to_device
+from forge.env import TORCHSTORE_USE_RDMA
 from forge.observability.metrics import record_metric, Reduce
 from forge.observability.perf_tracker import Tracer
 
@@ -111,7 +112,7 @@ class RLTrainer(ForgeActor):
     # Non JobConfig-related fields
     loss: Callable = lambda logits, **targets: logits
     state_dict_key: str = "model_state_dict"
-    use_dcp: bool = True
+    use_dcp: bool = not TORCHSTORE_USE_RDMA.get_value()
     dcp_path: str = "forge_dcp_tmp"
 
     def __post_init__(self):

src/forge/actors/trainer.py

@@ -102,13 +102,13 @@ def get_value(self) -> Any:
 
 MONARCH_HOSTMESH_V1 = EnvVar(
     name="MONARCH_HOSTMESH_V1",
-    default=True,
+    default=False,
     description="Whether or not to use Monarch's experimental hostmesh v1 APIs",
 )
 
 TORCHSTORE_USE_RDMA = EnvVar(
     name="TORCHSTORE_RDMA_ENABLED",
-    default=True,
+    default=False,
     description="Whether or not to use RDMA in TorchStore.",
 )
 

src/forge/env.py

@@ -8,9 +8,9 @@
 import logging
 from typing import Any, Union
 
-from monarch.actor import Actor, endpoint, get_or_spawn_controller, ProcMesh, this_proc
+from monarch.actor import Actor, endpoint, ProcMesh
 
-from forge.env import FORGE_DISABLE_METRICS
+from forge.env import FORGE_DISABLE_METRICS, MONARCH_HOSTMESH_V1
 from forge.observability.metrics import (
     BackendRole,
     get_logger_backend_class,
@@ -19,6 +19,13 @@
     reduce_metrics_states,
 )
 
+if MONARCH_HOSTMESH_V1.get_value():
+    from monarch._src.actor.v1.host_mesh import this_proc
+    from monarch._src.actor.v1.proc_mesh import get_or_spawn_controller
+else:
+    from monarch.actor import get_or_spawn_controller, this_proc
+
+
 logger = logging.getLogger(__name__)
 
 _global_logger = None