diff --git a/python/sglang/srt/disaggregation/cp_per_layer_transfer.py b/python/sglang/srt/disaggregation/cp_per_layer_transfer.py
index ec3c87e84..2c0155903 100644
--- a/python/sglang/srt/disaggregation/cp_per_layer_transfer.py
+++ b/python/sglang/srt/disaggregation/cp_per_layer_transfer.py
@@ -81,6 +81,7 @@ class PerLayerTransferContext:
         self.num_layers = num_layers
         self._batch_ids: List[int] = []
         self._submitted_layers: set[int] = set()
+        self._enqueued_layers: set[int] = set()
         self._enqueued = 0
         self._processed = 0
         self._failed = False
@@ -91,11 +92,19 @@ class PerLayerTransferContext:
         with self._cond:
             return self._failed
 
-    def note_enqueued(self) -> None:
-        """Called by the manager when it enqueues a (ctx, layer) item — the true count
-        of notifier fires for this request, which finish() waits on."""
+    def note_enqueued(self, layer_id: int) -> bool:
+        """Reserve layer_id for enqueue. Returns True only the FIRST time this layer is
+        seen, so the manager enqueues + counts each layer ONCE. Critical: the notifier
+        re-fires every layer on every subsequent forward while this ctx is still active
+        (awaiting finish); without dedup, _enqueued would grow unbounded (target>>layers)
+        and finish() could never reach it -> 30s timeouts. The first fire for a layer is
+        from THIS request's own forward, so its event is the correct one."""
         with self._cond:
+            if layer_id in self._enqueued_layers:
+                return False
+            self._enqueued_layers.add(layer_id)
             self._enqueued += 1
+            return True
 
     def _submit_one(self, layer_id: int) -> None:
         """Submit layer_id's transfer (KV assumed written). Updates state; logs +
@@ -222,6 +231,8 @@ class PerLayerTransferManager:
 
     def register(self, room, ctx: PerLayerTransferContext) -> None:
         with self._active_lock:
+            if room in self._active:
+                return  # already registered for its forward; don't overwrite (leak)
             self._active[room] = ctx
 
     def has_active(self) -> bool:
@@ -242,8 +253,8 @@ class PerLayerTransferManager:
             else:
                 event.record()
         for ctx in ctxs:
-            ctx.note_enqueued()
-            self._q.put((ctx, layer_id, event))
+            if ctx.note_enqueued(layer_id):  # enqueue each layer ONCE per ctx
+                self._q.put((ctx, layer_id, event))
 
     def has_room(self, room) -> bool:
         with self._active_lock:
diff --git a/test/registered/unit/disaggregation/test_cp_per_layer_transfer.py b/test/registered/unit/disaggregation/test_cp_per_layer_transfer.py
index 73e6a503b..d9e2eb31a 100644
--- a/test/registered/unit/disaggregation/test_cp_per_layer_transfer.py
+++ b/test/registered/unit/disaggregation/test_cp_per_layer_transfer.py
@@ -102,6 +102,20 @@ class TestPerLayerTransferContext(unittest.TestCase):
         ctx.submit_layer(0, None)
         self.assertEqual(ctx.finish(), -9)
 
+    def test_repeated_enqueues_dedup_so_finish_cannot_hang(self):
+        # the notifier re-fires every layer on every subsequent forward while the ctx
+        # is still active; note_enqueued must dedup so target stays capped at the layer
+        # count and finish() completes (the high-cache-hit hang was target>>layers).
+        eng = _FakeEngine()
+        ctx = _ctx(eng, num_layers=3)
+        for _forward in range(5):  # 5 forwards all re-fire layers 0,1,2
+            for L in range(3):
+                self.assertEqual(ctx.note_enqueued(L), _forward == 0)  # True only 1st
+                if _forward == 0:
+                    ctx.submit_layer(L, None)
+        self.assertEqual(len(eng.submits), 3)  # 3 unique despite 15 note_enqueued
+        self.assertEqual(ctx.finish(timeout=0.5), 0)  # completes, no 30s timeout
+
     def test_finish_falls_back_for_notifier_missed_layers(self):
         # only layer 0 fired by the notifier; finish must SYNCHRONOUSLY transfer the
         # missing layers 1,2 (e.g. an MTP buffer not hooked) -> all 3 moved, success.
@@ -155,8 +169,12 @@ class _MockCtx:
         self.finished = False
         self.failed = False
 
-    def note_enqueued(self):
+    def note_enqueued(self, layer_id):
+        if layer_id in getattr(self, "_enq_layers", set()):
+            return False
+        self._enq_layers = getattr(self, "_enq_layers", set()) | {layer_id}
         self.enqueued += 1
+        return True
 
     def submit_layer(self, layer_id, event):
         self.submitted.append((layer_id, event))