diff --git a/miles/backends/megatron_utils/actor.py b/miles/backends/megatron_utils/actor.py
index 9d93db11994..e43a4500178 100644
--- a/miles/backends/megatron_utils/actor.py
+++ b/miles/backends/megatron_utils/actor.py
@@ -613,7 +613,7 @@ def _ensure_cpu_bucket_cache(self):
         if not hasattr(self, "_cpu_bucket_cache") or self._cpu_bucket_cache is None:
             from .update_weight.cpu_bucket_cache import CPUBucketCache
 
-            max_bytes = int(getattr(self.args, "miles_model_update_bucket_size_mb", 512)) * 1024 * 1024
+            max_bytes = int(self.args.miles_model_update_bucket_size_mb) * 1024 * 1024
             self._cpu_bucket_cache = CPUBucketCache(max_bucket_size_bytes=max_bytes)
             # F20: bucket build / sync session each acquire this lock for
             # the whole critical section (single-method-single-critical-
diff --git a/miles/utils/rlix_validation.py b/miles/utils/rlix_validation.py
index 5d6f7727344..48b636fa3a8 100644
--- a/miles/utils/rlix_validation.py
+++ b/miles/utils/rlix_validation.py
@@ -335,9 +335,7 @@ def assert_rlix_topology(args: Any, sglang_config: Any | None = None) -> None:
         )
 
     # --- S2: bucket size <= post-wake free VRAM (NCCL broadcast / cuda_ipc paths)
-    bucket_size_bytes = (
-        int(getattr(args, "miles_model_update_bucket_size_mb", 512) or 512) * 1024 * 1024
-    )
+    bucket_size_bytes = int(args.miles_model_update_bucket_size_mb) * 1024 * 1024
     transport = getattr(args, "model_update_transport", "cuda_ipc")
     has_gpu_staging = _topology_has_non_colocate_engines(args) or transport == "cuda_ipc"
     if has_gpu_staging: