diff --git a/nemo_rl/algorithms/grpo_sync.py b/nemo_rl/algorithms/grpo_sync.py
index 4ece3f81da..816e7e6d83 100644
--- a/nemo_rl/algorithms/grpo_sync.py
+++ b/nemo_rl/algorithms/grpo_sync.py
@@ -76,6 +76,7 @@
 from nemo_rl.environments.interfaces import EnvironmentInterface
 from nemo_rl.experience.sync_rollout_actor import SyncRolloutActor
 from nemo_rl.models.generation.interfaces import GenerationInterface
+from nemo_rl.models.generation.megatron import MegatronGeneration
 from nemo_rl.models.policy.interfaces import ColocatablePolicyInterface
 from nemo_rl.utils.checkpoint import CheckpointManager
 from nemo_rl.utils.logger import Logger, print_message_log_samples
@@ -404,7 +405,10 @@ def grpo_train_sync(
 
     kv_scales_cache = None  # Cache reused for computed kv scales
 
-    NEED_REFIT = True
+    NEED_REFIT = not (
+        isinstance(policy_generation, MegatronGeneration)
+        and master_config.policy["generation"]["colocated"]["enabled"]
+    )
     # If policy_generation is None, use the policy as the generation interface (megatron framework backend)
     if policy_generation is None:
         policy_generation = policy  # type: ignore