diff --git a/examples/on_policy_distillation/agentic_opd/alfworld/run-alfworld-opd-qwen35-35B-A3B-8xgpu.sh b/examples/on_policy_distillation/agentic_opd/alfworld/run-alfworld-opd-qwen35-35B-A3B-8xgpu.sh
index c03272149..d61aa9e5b 100644
--- a/examples/on_policy_distillation/agentic_opd/alfworld/run-alfworld-opd-qwen35-35B-A3B-8xgpu.sh
+++ b/examples/on_policy_distillation/agentic_opd/alfworld/run-alfworld-opd-qwen35-35B-A3B-8xgpu.sh
@@ -157,7 +157,6 @@ SGLANG_ARGS=(
    --rollout-num-gpus-per-engine ${ROLLOUT_NUM_GPUS_PER_ENGINE:-4}
    --sglang-mem-fraction-static ${STUDENT_MEM_FRACTION:-0.7}
    --sglang-load-format dummy
-   --sglang-enable-weights-cpu-backup
    --sglang-max-running-requests 64
 )
 
diff --git a/examples/on_policy_distillation/agentic_opd/webshop/run-webshop-opd-qwen35-35B-A3B-8xgpu.sh b/examples/on_policy_distillation/agentic_opd/webshop/run-webshop-opd-qwen35-35B-A3B-8xgpu.sh
index 2e1ffc7f6..f04dae9cc 100644
--- a/examples/on_policy_distillation/agentic_opd/webshop/run-webshop-opd-qwen35-35B-A3B-8xgpu.sh
+++ b/examples/on_policy_distillation/agentic_opd/webshop/run-webshop-opd-qwen35-35B-A3B-8xgpu.sh
@@ -165,7 +165,6 @@ SGLANG_ARGS=(
    --rollout-num-gpus-per-engine ${ROLLOUT_NUM_GPUS_PER_ENGINE:-4}
    --sglang-mem-fraction-static ${STUDENT_MEM_FRACTION:-0.7}
    --sglang-load-format dummy
-   --sglang-enable-weights-cpu-backup
    --sglang-max-running-requests 64
 )
 
diff --git a/examples/on_policy_distillation/math_opd/run-opd-qwen35-35B-A3B-8xgpu-colocate.sh b/examples/on_policy_distillation/math_opd/run-opd-qwen35-35B-A3B-8xgpu-colocate.sh
index 90af97a62..759ee8527 100644
--- a/examples/on_policy_distillation/math_opd/run-opd-qwen35-35B-A3B-8xgpu-colocate.sh
+++ b/examples/on_policy_distillation/math_opd/run-opd-qwen35-35B-A3B-8xgpu-colocate.sh
@@ -133,7 +133,6 @@ SGLANG_ARGS=(
    --rollout-num-gpus-per-engine 4
    --sglang-mem-fraction-static 0.6
    --sglang-load-format dummy
-   --sglang-enable-weights-cpu-backup
    --sglang-max-running-requests 64
 )
 
diff --git a/examples/on_policy_distillation/mopd/run-mopd-qwen3-vl-2b-8xgpu-colocate.sh b/examples/on_policy_distillation/mopd/run-mopd-qwen3-vl-2b-8xgpu-colocate.sh
index e114d6756..ffd68f3b1 100755
--- a/examples/on_policy_distillation/mopd/run-mopd-qwen3-vl-2b-8xgpu-colocate.sh
+++ b/examples/on_policy_distillation/mopd/run-mopd-qwen3-vl-2b-8xgpu-colocate.sh
@@ -163,7 +163,6 @@ SGLANG_ARGS=(
    --rollout-num-gpus-per-engine 1
    --sglang-mem-fraction-static 0.8
    --sglang-load-format dummy
-   --sglang-enable-weights-cpu-backup
 )
 
 RESOURCE_JSON="{\"actor\": [1, ${ACTOR_GPUS}], \"rollout\": [1, ${ROLLOUT_GPUS}], \"teacher\": [1, ${TEACHER_GPUS}]}"
diff --git a/examples/on_policy_distillation/mopd/run-mopd-qwen35-35ba3b-16xgpu-colocate.sh b/examples/on_policy_distillation/mopd/run-mopd-qwen35-35ba3b-16xgpu-colocate.sh
index 5e9ff2f53..91cd83616 100644
--- a/examples/on_policy_distillation/mopd/run-mopd-qwen35-35ba3b-16xgpu-colocate.sh
+++ b/examples/on_policy_distillation/mopd/run-mopd-qwen35-35ba3b-16xgpu-colocate.sh
@@ -176,7 +176,6 @@ SGLANG_ARGS=(
    --sglang-mem-fraction-static 0.6
    --sglang-max-running-requests 128
    --sglang-load-format dummy
-   --sglang-enable-weights-cpu-backup
 )
 
 PARTIAL_ROLLOUT_ARGS=(
diff --git a/examples/on_policy_distillation/mopd/run-mopd-qwen35-9b-8xgpu-colocate.sh b/examples/on_policy_distillation/mopd/run-mopd-qwen35-9b-8xgpu-colocate.sh
index 660a9ff02..f2648a60d 100755
--- a/examples/on_policy_distillation/mopd/run-mopd-qwen35-9b-8xgpu-colocate.sh
+++ b/examples/on_policy_distillation/mopd/run-mopd-qwen35-9b-8xgpu-colocate.sh
@@ -163,7 +163,6 @@ SGLANG_ARGS=(
    --rollout-num-gpus-per-engine 2
    --sglang-mem-fraction-static 0.7
    --sglang-load-format dummy
-   --sglang-enable-weights-cpu-backup
 )
 
 RESOURCE_JSON="{\"actor\": [1, ${ACTOR_GPUS}], \"rollout\": [1, ${ROLLOUT_GPUS}], \"teacher\": [1, ${TEACHER_GPUS}]}"
diff --git a/examples/on_policy_distillation/vision_opd/run-opd-qwen3.5_35ba3b-122ba10b-128xgpu-colocate.sh b/examples/on_policy_distillation/vision_opd/run-opd-qwen3.5_35ba3b-122ba10b-128xgpu-colocate.sh
index 3164c65de..73e4c3a7f 100644
--- a/examples/on_policy_distillation/vision_opd/run-opd-qwen3.5_35ba3b-122ba10b-128xgpu-colocate.sh
+++ b/examples/on_policy_distillation/vision_opd/run-opd-qwen3.5_35ba3b-122ba10b-128xgpu-colocate.sh
@@ -199,7 +199,6 @@ SGLANG_ARGS=(
    --sglang-mem-fraction-static ${STUDENT_MEM_FRACTION:-0.7}
    --sglang-load-format dummy
    --sglang-cuda-graph-bs 1 2 4 8 $(seq 16 8 256)
-   --sglang-enable-weights-cpu-backup
 )
 
 WANDB_ARGS=(
diff --git a/examples/on_policy_distillation/vision_opd/run-vision-opd-qwen3.5_9b-35ba3b-8xgpu-2teacher-colocate.sh b/examples/on_policy_distillation/vision_opd/run-vision-opd-qwen3.5_9b-35ba3b-8xgpu-2teacher-colocate.sh
index 8e7afbe08..6d19bff7a 100755
--- a/examples/on_policy_distillation/vision_opd/run-vision-opd-qwen3.5_9b-35ba3b-8xgpu-2teacher-colocate.sh
+++ b/examples/on_policy_distillation/vision_opd/run-vision-opd-qwen3.5_9b-35ba3b-8xgpu-2teacher-colocate.sh
@@ -129,7 +129,6 @@ SGLANG_ARGS=(
    --rollout-num-gpus-per-engine "${ROLLOUT_NUM_GPUS_PER_ENGINE:-2}"
    --sglang-mem-fraction-static "${STUDENT_MEM_FRACTION:-0.8}"
    --sglang-load-format dummy
-   --sglang-enable-weights-cpu-backup
    --sglang-cuda-graph-bs 1 2 4 8 $(seq 16 8 256)
 )
 
diff --git a/examples/on_policy_distillation/vision_opd/run-vision-opd-qwen3.5_9b-35ba3b-8xgpu-colocate.sh b/examples/on_policy_distillation/vision_opd/run-vision-opd-qwen3.5_9b-35ba3b-8xgpu-colocate.sh
index 50de1e445..ec3b7f174 100755
--- a/examples/on_policy_distillation/vision_opd/run-vision-opd-qwen3.5_9b-35ba3b-8xgpu-colocate.sh
+++ b/examples/on_policy_distillation/vision_opd/run-vision-opd-qwen3.5_9b-35ba3b-8xgpu-colocate.sh
@@ -133,7 +133,6 @@ SGLANG_ARGS=(
    --rollout-num-gpus-per-engine "${ROLLOUT_NUM_GPUS_PER_ENGINE:-2}"
    --sglang-mem-fraction-static "${STUDENT_MEM_FRACTION:-0.8}"
    --sglang-load-format dummy
-   --sglang-enable-weights-cpu-backup
    --sglang-cuda-graph-bs 1 2 4 8 $(seq 16 8 256)
 )
 
diff --git a/relax/utils/arguments.py b/relax/utils/arguments.py
index f7d1e710e..7b084569e 100644
--- a/relax/utils/arguments.py
+++ b/relax/utils/arguments.py
@@ -18,6 +18,7 @@
 from relax.utils.opd.opd_utils import (
     add_opd_arguments,
     is_managed_opd_teacher_enabled,
+    maybe_enable_sglang_weights_cpu_backup,
     teacher_sglang_parse_args,
     validate_managed_opd_teacher_colocate_args,
     validate_opd_args,
@@ -3455,6 +3456,7 @@ def slime_validate_args(args):
         args.offload_train = False
     if args.offload_rollout is None:
         args.offload_rollout = False
+    maybe_enable_sglang_weights_cpu_backup(args)
 
     if args.use_critic:
         args.offload_train = True
diff --git a/relax/utils/opd/opd_utils.py b/relax/utils/opd/opd_utils.py
index ce052263a..6ab08d0c8 100644
--- a/relax/utils/opd/opd_utils.py
+++ b/relax/utils/opd/opd_utils.py
@@ -143,6 +143,9 @@ def build_teacher_overrides(args: Any, colocate_sync: bool = False) -> dict[str,
     overrides["model_path"] = args.teacher_hf_checkpoint
     overrides.setdefault("load_format", "auto")
     overrides.setdefault("enable_memory_saver", colocate_sync)
+    if overrides.get("enable_memory_saver"):
+        # Release/resume without a CPU backup leaves weight pages uninitialized (uniform output).
+        overrides.setdefault("enable_weights_cpu_backup", True)
     return overrides
 
 
@@ -460,6 +463,18 @@ def validate_managed_opd_teacher_colocate_args(args: Any) -> None:
         )
 
 
+def maybe_enable_sglang_weights_cpu_backup(args: Any) -> None:
+    # Rollout sleep/wake without a CPU backup leaves weight pages uninitialized.
+    if not getattr(args, "use_opd", False):
+        return
+    if not getattr(args, "offload_rollout", False):
+        return
+    if getattr(args, "sglang_enable_weights_cpu_backup", False):
+        return
+    logger.info("OPD with offloaded rollout: auto-enabling --sglang-enable-weights-cpu-backup.")
+    args.sglang_enable_weights_cpu_backup = True
+
+
 def add_opd_arguments(parser: Any) -> Any:
     parser.add_argument(
         "--use-opd",
diff --git a/tests/distributed/ray/test_teacher_sglang_overrides.py b/tests/distributed/ray/test_teacher_sglang_overrides.py
index ef2ff6a8d..d07a5c376 100644
--- a/tests/distributed/ray/test_teacher_sglang_overrides.py
+++ b/tests/distributed/ray/test_teacher_sglang_overrides.py
@@ -6,6 +6,7 @@
 from relax.utils.opd.opd_utils import (
     build_teacher_engine_args,
     build_teacher_overrides,
+    maybe_enable_sglang_weights_cpu_backup,
     teacher_sglang_parse_args,
 )
 
@@ -52,3 +53,35 @@ def add_cli_args(parser):
 
     assert args.teacher_sglang_mem_fraction_static == 0.73
     assert not hasattr(args, "teacher_sglang_model_path")
+
+
+def test_opd_with_offloaded_rollout_enables_weights_cpu_backup():
+    args = SimpleNamespace(use_opd=True, offload_rollout=True)
+
+    maybe_enable_sglang_weights_cpu_backup(args)
+
+    assert args.sglang_enable_weights_cpu_backup is True
+
+
+def test_opd_without_offload_keeps_weights_cpu_backup_untouched():
+    args = SimpleNamespace(use_opd=True, offload_rollout=False)
+
+    maybe_enable_sglang_weights_cpu_backup(args)
+
+    assert not hasattr(args, "sglang_enable_weights_cpu_backup")
+
+
+def test_non_opd_keeps_weights_cpu_backup_untouched():
+    args = SimpleNamespace(offload_rollout=True)
+
+    maybe_enable_sglang_weights_cpu_backup(args)
+
+    assert not hasattr(args, "sglang_enable_weights_cpu_backup")
+
+
+def test_explicit_weights_cpu_backup_is_preserved():
+    args = SimpleNamespace(use_opd=True, offload_rollout=True, sglang_enable_weights_cpu_backup=True)
+
+    maybe_enable_sglang_weights_cpu_backup(args)
+
+    assert args.sglang_enable_weights_cpu_backup is True