update configs

Author: Felipe Mello

apps/grpo/main.py

@@ -334,7 +334,7 @@ async def main(cfg: DictConfig):
 
     # initialize before spawning services
     metric_logging_cfg = cfg.get("metric_logging", {"console": {"log_per_rank": False}})
-    mlogger = await get_or_create_metric_logger()
+    mlogger = await get_or_create_metric_logger(actor_name="Controller")
 
     # ---- Setup services ---- #
     await ts.initialize(strategy=ts.ControllerStorageVolumes())

apps/grpo/qwen3_1_7b.yaml

@@ -14,14 +14,15 @@ rollout_threads: 1   # Recommended to set equal to policy.num_replicas
 
 
 # Observability configuration
+# logging_mode: global_reduce, per_rank_reduce, per_rank_no_reduce
 metric_logging:
   wandb:
-    project: "test"
+    project: "grpo-training"
     group: "grpo_exp_${oc.env:USER}"
-    logging_mode: "per_rank_no_reduce"
-    "per_rank_share_run": True
+    logging_mode: "global_reduce"
+    per_rank_share_run: False
   console:
-    logging_mode: "per_rank_no_reduce"
+    logging_mode: "global_reduce"
 
 # Dataset configuration
 dataset:

apps/grpo/qwen3_32b.yaml

@@ -14,13 +14,15 @@ off_by_n: 1 # Off by one by default
 rollout_threads: 1   # Recommended to set equal to policy.num_replicas
 
 # Observability configuration
+# logging_mode: global_reduce, per_rank_reduce, per_rank_no_reduce
 metric_logging:
   wandb:
     project: "grpo-training"
     group: "grpo_exp_${oc.env:USER}"
-    reduce_across_ranks: True
+    logging_mode: "global_reduce"
+    per_rank_share_run: False
   console:
-    reduce_across_ranks: True
+    logging_mode: "global_reduce"
 
 # Dataset configuration
 dataset:

apps/grpo/qwen3_8b.yaml

@@ -10,13 +10,15 @@ model: "Qwen/Qwen3-8B"
 off_by_n: 1 # Off by one by default
 
 # Observability configuration
+# logging_mode: global_reduce, per_rank_reduce, per_rank_no_reduce
 metric_logging:
   wandb:
     project: "grpo-training"
     group: "grpo_exp_${oc.env:USER}"
-    reduce_across_ranks: True
+    logging_mode: "global_reduce"
+    per_rank_share_run: False
   console:
-    reduce_across_ranks: True
+    logging_mode: "global_reduce"
 
 # Dataset configuration
 dataset:

apps/vllm/main.py

@@ -27,8 +27,7 @@
 
 async def run(cfg: DictConfig):
     metric_logging_cfg = cfg.get("metric_logging", {"console": {"log_per_rank": False}})
-    mlogger = await get_or_create_metric_logger()
-    await mlogger.init_backends.call_one(metric_logging_cfg)
+    mlogger = await get_or_create_metric_logger(actor_name="Controller")
 
     if (prompt := cfg.get("prompt")) is None:
         gd = cfg.policy.get("sampling_config", {}).get("guided_decoding", False)
@@ -37,6 +36,9 @@ async def run(cfg: DictConfig):
     print("Spawning service...")
     policy = await Policy.options(**cfg.services.policy).as_service(**cfg.policy)
 
+    # initialize after spawning services
+    await mlogger.init_backends.call_one(metric_logging_cfg)
+
     import time
 
     print("Requesting generation...")

src/forge/observability/metrics.py

@@ -15,10 +15,10 @@
 
 import pytz
 
-from monarch.actor import current_rank
-
 from forge.observability.utils import get_actor_name_with_rank
 
+from monarch.actor import current_rank
+
 logger = logging.getLogger(__name__)