diff --git a/applications/ColossalChat/coati/distributed/consumer.py b/applications/ColossalChat/coati/distributed/consumer.py
index e360392e7..ba7d882c9 100644
--- a/applications/ColossalChat/coati/distributed/consumer.py
+++ b/applications/ColossalChat/coati/distributed/consumer.py
@@ -365,7 +365,7 @@ class SimpleConsumer(BaseConsumer):
         self.model = AutoModelForCausalLM.from_pretrained(path, **model_config)
         self.model.train()
         self.model.gradient_checkpointing_enable()
-        self.optimizer = HybridAdam(self.model.parameters(), lr=1e-3)
+        self.optimizer = HybridAdam(self.model.parameters(), lr=1e-3, weight_decay=0.01)
         self.accum_loss = torch.zeros(1, device=self.device)
 
     def setup(self):
diff --git a/applications/ColossalChat/coati/distributed/grpo_consumer.py b/applications/ColossalChat/coati/distributed/grpo_consumer.py
index a3f1a1cbb..424d46098 100644
--- a/applications/ColossalChat/coati/distributed/grpo_consumer.py
+++ b/applications/ColossalChat/coati/distributed/grpo_consumer.py
@@ -72,7 +72,11 @@ class GRPOConsumer(BaseConsumer):
         self.policy_model = AutoModelForCausalLM.from_pretrained(path, **model_config)
         self.policy_model.train()
         self.policy_model.gradient_checkpointing_enable()
-        self.optimizer = HybridAdam(self.policy_model.parameters(), lr=grpo_config.get("lr", 1e-6))
+        self.optimizer = HybridAdam(
+            self.policy_model.parameters(),
+            lr=grpo_config.get("lr", 1e-6),
+            weight_decay=grpo_config.get("weight_decay", 0.01),
+        )
         self.accum_loss = torch.zeros(1, device=self.device)
         self.accum_kl = torch.zeros(1, device=self.device)
         self.accum_entropy = torch.zeros(1, device=self.device)