Merge remote-tracking branch 'upstream/main' into cleanup-selective-log-sm

Author: joecummings

.github/workflows/docs.yml

@@ -65,7 +65,7 @@ jobs:
 
           # Verify dependencies can be imported before building docs
           echo "Verifying dependencies..."
-          python -c "import forge; print('✓ forge imported successfully')"
+          python -c "import forge; print('✓ torchforge imported successfully')"
           python -c "import monarch; print('✓ monarch imported successfully')"
 
           # Build docs with -W (warnings as errors) and --keep-going to see all issues
@@ -112,15 +112,15 @@ jobs:
           s3-bucket: doc-previews
           if-no-files-found: error
           path: docs
-          s3-prefix: meta-pytorch/forge/${{ github.event.pull_request.number }}
+          s3-prefix: meta-pytorch/torchforge/${{ github.event.pull_request.number }}
 
   upload:
     runs-on: ubuntu-latest
     permissions:
       # Grant write permission here so that the doc can be pushed to gh-pages branch
       contents: write
     needs: build-docs
-    if: github.repository == 'meta-pytorch/forge' && github.event_name == 'push' && (github.ref == 'refs/heads/main' || startsWith(github.ref, 'refs/tags/v') || github.event_name == 'workflow_dispatch')
+    if: github.repository == 'meta-pytorch/torchforge' && github.event_name == 'push' && (github.ref == 'refs/heads/main' || startsWith(github.ref, 'refs/tags/v') || github.event_name == 'workflow_dispatch')
     steps:
       - name: Checkout
         uses: actions/checkout@v4

.meta/mast/qwen3_14b_mast.yaml

@@ -19,9 +19,9 @@ metric_logging:
   wandb:
     project: "grpo-training"
     group: "grpo_exp_${oc.env:USER}"
-    reduce_across_ranks: True
+    logging_mode: global_reduce
   console:
-    reduce_across_ranks: True
+    logging_mode: global_reduce
 
 # Dataset configuration
 dataset:

.meta/mast/qwen3_1_7b_mast.yaml

@@ -19,9 +19,9 @@ metric_logging:
   wandb:
     project: "grpo-training"
     group: "grpo_exp_${oc.env:USER}"
-    reduce_across_ranks: True
+    logging_mode: global_reduce
   console:
-    reduce_across_ranks: True
+    logging_mode: global_reduce
 
 # Dataset configuration
 dataset:

.meta/mast/qwen3_32b_mast.yaml

@@ -19,9 +19,9 @@ metric_logging:
   wandb:
     project: "grpo-training"
     group: "grpo_exp_${oc.env:USER}"
-    reduce_across_ranks: True
+    logging_mode: global_reduce
   console:
-    reduce_across_ranks: True
+    logging_mode: global_reduce
 
 # Dataset configuration
 dataset:

.meta/mast/qwen3_4b_mast.yaml

@@ -19,9 +19,9 @@ metric_logging:
   wandb:
     project: "grpo-training"
     group: "grpo_exp_${oc.env:USER}"
-    reduce_across_ranks: True
+    logging_mode: global_reduce
   console:
-    reduce_across_ranks: True
+    logging_mode: global_reduce
 
 # Dataset configuration
 dataset:

.meta/mast/qwen3_8b_mast.yaml

@@ -19,9 +19,9 @@ metric_logging:
   wandb:
     project: "grpo-training"
     group: "grpo_exp_${oc.env:USER}"
-    reduce_across_ranks: True
+    logging_mode: global_reduce
   console:
-    reduce_across_ranks: True
+    logging_mode: global_reduce
 
 # Dataset configuration
 dataset:

apps/grpo/main.py

@@ -304,7 +304,7 @@ async def main(cfg: DictConfig):
     else:
         provisioner = await init_provisioner()
 
-    metric_logging_cfg = cfg.get("metric_logging", {"console": {"log_per_rank": False}})
+    metric_logging_cfg = cfg.get("metric_logging", {})
     mlogger = await get_or_create_metric_logger(process_name="Controller")
     await mlogger.init_backends.call_one(metric_logging_cfg)
 

apps/grpo/qwen3_1_7b.yaml

@@ -16,11 +16,11 @@ rollout_threads: 1   # Recommended to set equal to policy.num_replicas
 # Observability configuration
 metric_logging:
   wandb:
-    project: "grpo-training"
-    group: "grpo_exp_${oc.env:USER}"
-    reduce_across_ranks: True
+    project: grpo-training
+    group: grpo_exp_${oc.env:USER}
+    logging_mode: global_reduce # global_reduce, per_rank_reduce, per_rank_no_reduce
   console:
-    reduce_across_ranks: True
+    logging_mode: global_reduce
 
 # Dataset configuration
 dataset:

apps/grpo/qwen3_32b.yaml

@@ -19,11 +19,11 @@ rollout_threads: 32 # make this 4x the number of policy replicas seems to work w
 # Observability configuration
 metric_logging:
   wandb:
-    project: "grpo-training"
-    group: "grpo_exp_${oc.env:USER}"
-    reduce_across_ranks: True
+    project: grpo-training
+    group: grpo_exp_${oc.env:USER}
+    logging_mode: global_reduce # global_reduce, per_rank_reduce, per_rank_no_reduce
   console:
-    reduce_across_ranks: True
+    logging_mode: global_reduce
 
 # Dataset configuration
 dataset:

apps/grpo/qwen3_8b.yaml

@@ -12,11 +12,11 @@ off_by_n: 1 # Off by one by default
 # Observability configuration
 metric_logging:
   wandb:
-    project: "grpo-training"
-    group: "grpo_exp_${oc.env:USER}"
-    reduce_across_ranks: True
+    project: grpo-training
+    group: grpo_exp_${oc.env:USER}
+    logging_mode: global_reduce # global_reduce, per_rank_reduce, per_rank_no_reduce
   console:
-    reduce_across_ranks: True
+    logging_mode: global_reduce
 
 # Dataset configuration
 dataset: