rollout loops

allenwang28 · allenwang28 · commit 061dbdcfc4a8 · 2025-10-10T12:22:27.000-07:00
diff --git a/apps/grpo/qwen3_32b.yaml b/apps/grpo/qwen3_32b.yaml
@@ -14,7 +14,7 @@ provisioner:
   launcher: slurm
 
 # Main loop configuration
-rollout_threads: 32 # make this equal to the batch size seems to work well
+rollout_threads: 32 # make this 4x the number of policy replicas seems to work well
 
 # Observability configuration
 metric_logging:
diff --git a/apps/mast/qwen3_32b_mast.yaml b/apps/mast/qwen3_32b_mast.yaml
@@ -13,7 +13,7 @@ job_name: forge-qwen3-32b
 checkpoint_folder: /mnt/wsfuse/teamforge/forge_runs/
 
 # Main loop configuration
-rollout_threads: 8
+rollout_threads: 32 # make this 4x the number of policy replicas seems to work well
 
 # Observability configuration
 metric_logging:
