diff --git a/verl/utils/torch_functional.py b/verl/utils/torch_functional.py
index 726ca1d5609..9754d989344 100644
--- a/verl/utils/torch_functional.py
+++ b/verl/utils/torch_functional.py
@@ -473,7 +473,9 @@ def get_constant_schedule_with_warmup(
     last_epoch: int = -1,
 ):
     def lr_lambda(current_step):
-        return min(1, float(current_step) / float(max(1, num_warmup_steps)))
+        if current_step < num_warmup_steps:
+            return float(current_step) / float(max(1.0, num_warmup_steps))
+        return 1.0
 
     return LambdaLR(optimizer, lr_lambda, last_epoch)
 
diff --git a/verl/workers/fsdp_workers.py b/verl/workers/fsdp_workers.py
index fffce53c062..99422f6fcfc 100644
--- a/verl/workers/fsdp_workers.py
+++ b/verl/workers/fsdp_workers.py
@@ -584,9 +584,9 @@ def update_actor(self, data: DataProto):
             metrics["perf/max_memory_reserved_gb"] = torch.cuda.max_memory_reserved() / (1024**3)
             metrics["perf/cpu_memory_used_gb"] = psutil.virtual_memory().used / (1024**3)
 
-            self.actor_lr_scheduler.step()
             lr = self.actor_lr_scheduler.get_last_lr()[0]
             metrics["actor/lr"] = lr
+            self.actor_lr_scheduler.step()
 
             # TODO: here, we should return all metrics
             output = DataProto(meta_info={"metrics": metrics})