Async by 1

Author: joecummings

apps/grpo/qwen3_1_7b.yaml

@@ -2,7 +2,7 @@
 
 # Global configuration
 group_size: 8
-batch_size: 8
+batch_size: 16
 max_req_tokens: 512
 max_res_tokens: 512
 model: "Qwen/Qwen3-1.7B"
@@ -48,7 +48,7 @@ trainer:
 # Replay buffer configuration
 replay_buffer:
   batch_size: ${batch_size}
-  max_policy_age: 0
+  max_policy_age: 1 # Async by 1
   dp_size: 1
   service:
     procs_per_replica: 1