2025-11-13 nightly release (6b65eb4)

Author: pytorchbot

.meta/mast/env_setup.sh

@@ -9,7 +9,7 @@
 # setup_forge_env.sh - Setup conda environment and install forge with mounting
 
 # Configuration
-CONDA_ENV_NAME="forge:41468b33a03eaf2bf5b44517f418028a"
+CONDA_ENV_NAME="forge:314c3548ae691f4aa2e49f1b1fad06b3"
 
 # Colors for output
 RED='\033[0;31m'

apps/sft/main.py

@@ -81,34 +81,8 @@ def __init__(self, config: DictConfig):
         self.gradient_accumulation_steps = 1  # Example value, adjust as needed
         self._rank = current_rank().rank
         self._size = math.prod(current_size().values())
-        self._init_dist()
         super().__init__(job_config)
 
-    def _init_dist(self):
-        """Initializes torch distributed.
-
-        torchrun normally hands this, but we need to do it ourselves
-        in monarch for now.
-
-        We should consider putting this into ForgeActor, but having this
-        be explicit for now.
-
-        """
-        env = {
-            "RANK": str(self._rank),
-            "LOCAL_RANK": str(self._rank),
-            "LOCAL_WORLD_SIZE": str(self._size),
-            "GROUP_RANK": str(self._size),
-            "GROUP_WORLD_SIZE": str(self._size),
-            "ROLE_RANK": str(self._rank),
-            "ROLE_WORLD_SIZE": str(self._size),
-            "ROLE_NAME": "rank",
-            "WORLD_SIZE": str(self._size),
-            "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True",
-        }
-        os.environ.update(env)
-        logger.info("env: {}".format(env))
-
     async def setup_metric_logger(self):
         """Initialization happens in the main process. Here we just retrieve it"""
         mlogger = await get_or_create_metric_logger()

src/forge/controller/launcher.py

@@ -15,20 +15,18 @@
 from typing import Any
 
 import monarch
-
 import torchx.specs as specs
-
-from forge.types import Launcher, LauncherConfig
 from monarch._rust_bindings.monarch_hyperactor.alloc import AllocConstraints
 from monarch._rust_bindings.monarch_hyperactor.channel import ChannelTransport
-
 from monarch._rust_bindings.monarch_hyperactor.config import configure
 from monarch._src.actor.allocator import RemoteAllocator, TorchXRemoteAllocInitializer
 from monarch.actor import Actor, endpoint, ProcMesh
 from monarch.tools import commands
 from monarch.tools.commands import create, info
 from monarch.tools.config import Config, Workspace
 
+from forge.types import Launcher, LauncherConfig
+
 _MAST_AVAILABLE = False
 
 try:
@@ -267,7 +265,7 @@ async def launch_mast_job(self):
 
     def add_additional_packages(self, packages: "Packages") -> "Packages":
         packages.add_package("oil.oilfs:stable")
-        packages.add_package("manifold.manifoldfs")
+        packages.add_package("manifold.manifoldfs:prod")
         return packages
 
     def build_appdef(self) -> specs.AppDef: