diff --git a/python/sglang/srt/sampling/penaltylib/frequency_penalty.py b/python/sglang/srt/sampling/penaltylib/frequency_penalty.py
index 893a1c377..63d838574 100644
--- a/python/sglang/srt/sampling/penaltylib/frequency_penalty.py
+++ b/python/sglang/srt/sampling/penaltylib/frequency_penalty.py
@@ -1,9 +1,6 @@
 import torch
 
-from sglang.srt.sampling.penaltylib.orchestrator import (
-    BatchedPenalizerOrchestrator,
-    _BatchedPenalizer,
-)
+from sglang.srt.sampling.penaltylib.orchestrator import _BatchedPenalizer
 
 
 class BatchedFrequencyPenalizer(_BatchedPenalizer):
@@ -11,10 +8,6 @@ class BatchedFrequencyPenalizer(_BatchedPenalizer):
     Frequency penalizer penalizes tokens based on their frequency in the output.
     """
 
-    def __init__(self, orchestrator: BatchedPenalizerOrchestrator):
-        self.orchestrator = orchestrator
-        self._is_prepared = False
-
     def _is_required(self) -> bool:
         return any(
             req.sampling_params.frequency_penalty != 0.0
@@ -63,3 +56,8 @@ class BatchedFrequencyPenalizer(_BatchedPenalizer):
             [self.cumulated_frequency_penalties, their.cumulated_frequency_penalties],
             dim=0,
         )
+
+    def _teardown(self) -> None:
+        for name in ("frequency_penalties", "cumulated_frequency_penalties"):
+            if hasattr(self, name):
+                delattr(self, name)
diff --git a/python/sglang/srt/sampling/penaltylib/min_new_tokens.py b/python/sglang/srt/sampling/penaltylib/min_new_tokens.py
index da06265d9..08f76e1f1 100644
--- a/python/sglang/srt/sampling/penaltylib/min_new_tokens.py
+++ b/python/sglang/srt/sampling/penaltylib/min_new_tokens.py
@@ -1,9 +1,6 @@
 import torch
 
-from sglang.srt.sampling.penaltylib.orchestrator import (
-    BatchedPenalizerOrchestrator,
-    _BatchedPenalizer,
-)
+from sglang.srt.sampling.penaltylib.orchestrator import _BatchedPenalizer
 
 
 class BatchedMinNewTokensPenalizer(_BatchedPenalizer):
@@ -11,10 +8,6 @@ class BatchedMinNewTokensPenalizer(_BatchedPenalizer):
     Min new tokens penalizer penalizes tokens based on the length of the output.
     """
 
-    def __init__(self, orchestrator: BatchedPenalizerOrchestrator):
-        self.orchestrator = orchestrator
-        self._is_prepared = False
-
     def _is_required(self) -> bool:
         return any(
             req.sampling_params.min_new_tokens > 0 for req in self.orchestrator.reqs()
@@ -92,3 +85,9 @@ class BatchedMinNewTokensPenalizer(_BatchedPenalizer):
         self.len_output_tokens = torch.cat(
             [self.len_output_tokens, their.len_output_tokens], dim=0
         )
+
+    # Explicit resource cleanup to aid GC and free CUDA memory promptly
+    def _teardown(self) -> None:
+        for name in ("min_new_tokens", "stop_token_penalties", "len_output_tokens"):
+            if hasattr(self, name):
+                delattr(self, name)
diff --git a/python/sglang/srt/sampling/penaltylib/orchestrator.py b/python/sglang/srt/sampling/penaltylib/orchestrator.py
index 1abd255cb..7ef123f55 100644
--- a/python/sglang/srt/sampling/penaltylib/orchestrator.py
+++ b/python/sglang/srt/sampling/penaltylib/orchestrator.py
@@ -77,9 +77,8 @@ class BatchedPenalizerOrchestrator:
             return
 
         if len(keep_indices) == 0:
-            self.is_required = False
-            for penalizer in self.penalizers.values():
-                penalizer.teardown()
+            # No requests left in the batch, fully release orchestrator resources
+            self.release()
             return
 
         is_required = False
@@ -92,6 +91,23 @@ class BatchedPenalizerOrchestrator:
                 penalizer.teardown()
         self.is_required = is_required
 
+    # Resource management helpers
+    def release(self) -> None:
+        """Release all penalizers and break references so GC can reclaim promptly."""
+        for penalizer in self.penalizers.values():
+            penalizer.teardown()
+        self.penalizers.clear()
+        # Break reference to ScheduleBatch
+        self._batch_ref = None
+        self.is_required = False
+
+    # Context manager support
+    def __enter__(self) -> "BatchedPenalizerOrchestrator":
+        return self
+
+    def __exit__(self, exc_type, exc, tb) -> None:
+        self.release()
+
     def merge(self, their: "BatchedPenalizerOrchestrator"):
         """
         Merge the penalizers of another orchestrator into this one.
@@ -116,6 +132,22 @@ class _BatchedPenalizer(abc.ABC):
     An abstract class for a batched penalizer.
     """
 
+    def __init__(self, orchestrator: BatchedPenalizerOrchestrator):
+        self._orchestrator_ref: weakref.ReferenceType[BatchedPenalizerOrchestrator] = (
+            weakref.ref(orchestrator)
+        )
+        self._is_prepared = False
+
+    @property
+    def orchestrator(self) -> BatchedPenalizerOrchestrator:
+        orch: Optional[BatchedPenalizerOrchestrator] = self._orchestrator_ref()
+        # This should never happen, but we need to handle it gracefully
+        if orch is None:
+            raise RuntimeError(
+                "BatchedPenalizerOrchestrator has been garbage-collected"
+            )
+        return orch
+
     def is_prepared(self) -> bool:
         return self._is_prepared
 
@@ -135,6 +167,7 @@ class _BatchedPenalizer(abc.ABC):
             return False
 
     def teardown(self):
+        self._teardown()
         self._is_prepared = False
 
     def cumulate_output_tokens(self, output_ids: torch.Tensor):
@@ -207,3 +240,10 @@ class _BatchedPenalizer(abc.ABC):
         Merge the penalizer with another penalizer.
         """
         pass
+
+    @abc.abstractmethod
+    def _teardown(self):
+        """
+        Teardown the penalizer.
+        """
+        pass
diff --git a/python/sglang/srt/sampling/penaltylib/presence_penalty.py b/python/sglang/srt/sampling/penaltylib/presence_penalty.py
index 4f3a6ace3..1c045039e 100644
--- a/python/sglang/srt/sampling/penaltylib/presence_penalty.py
+++ b/python/sglang/srt/sampling/penaltylib/presence_penalty.py
@@ -1,9 +1,6 @@
 import torch
 
-from sglang.srt.sampling.penaltylib.orchestrator import (
-    BatchedPenalizerOrchestrator,
-    _BatchedPenalizer,
-)
+from sglang.srt.sampling.penaltylib.orchestrator import _BatchedPenalizer
 
 
 class BatchedPresencePenalizer(_BatchedPenalizer):
@@ -11,10 +8,6 @@ class BatchedPresencePenalizer(_BatchedPenalizer):
     Presence penalizer penalizes tokens based on their presence in the output.
     """
 
-    def __init__(self, orchestrator: BatchedPenalizerOrchestrator):
-        self.orchestrator = orchestrator
-        self._is_prepared = False
-
     def _is_required(self) -> bool:
         return any(
             req.sampling_params.presence_penalty != 0.0
@@ -63,3 +56,8 @@ class BatchedPresencePenalizer(_BatchedPenalizer):
             [self.cumulated_presence_penalties, their.cumulated_presence_penalties],
             dim=0,
         )
+
+    def _teardown(self) -> None:
+        for name in ("presence_penalties", "cumulated_presence_penalties"):
+            if hasattr(self, name):
+                delattr(self, name)