From ba97cc6397ac98b0d889609598cc18ad365d462c Mon Sep 17 00:00:00 2001
From: Ke Bao <ispobaoke@gmail.com>
Date: Fri, 21 Aug 2026 01:00:50 +0800
Subject: [PATCH] Skip empty linear-attention state buffers in PD transfer
 (#35689)

---
 python/sglang/srt/mem_cache/memory_pool.py    |  4 ++
 .../test_mamba_state_transfer_buffers.py      | 59 +++++++++++++++++++
 2 files changed, 63 insertions(+)
 create mode 100644 test/registered/unit/mem_cache/test_mamba_state_transfer_buffers.py

diff --git a/python/sglang/srt/mem_cache/memory_pool.py b/python/sglang/srt/mem_cache/memory_pool.py
index 420fa25b8..1bbc997d9 100644
--- a/python/sglang/srt/mem_cache/memory_pool.py
+++ b/python/sglang/srt/mem_cache/memory_pool.py
@@ -1074,6 +1074,10 @@ class MambaPool:
             tensors = value if isinstance(value, list) else [value]
             slice_axis = self.conv_slice_axis if field == "conv" else 0
             for state_tensor in tensors:
+                # A ShortConv layer has no temporal state, so that buffer is
+                # empty. Advertising it fails the whole batch registration.
+                if state_tensor.numel() == 0:
+                    continue
                 yield field, state_tensor, slice_axis
 
     def get_contiguous_buf_infos(self):
diff --git a/test/registered/unit/mem_cache/test_mamba_state_transfer_buffers.py b/test/registered/unit/mem_cache/test_mamba_state_transfer_buffers.py
new file mode 100644
index 000000000..12cb8b4bb
--- /dev/null
+++ b/test/registered/unit/mem_cache/test_mamba_state_transfer_buffers.py
@@ -0,0 +1,59 @@
+import unittest
+
+import torch
+
+from sglang.srt.mem_cache.memory_pool import MambaPool
+from sglang.test.ci.ci_register import register_cpu_ci
+
+register_cpu_ci(est_time=5, suite="base-a-test-cpu")
+
+NUM_LAYERS = 2
+NUM_SLOTS = 3
+
+
+def _pool(temporal: torch.Tensor, num_conv: int = 2) -> MambaPool:
+    """A MambaPool stub carrying only what the transfer accessors read."""
+    pool = object.__new__(MambaPool)
+    pool.num_mamba_layers = NUM_LAYERS
+    pool.conv_slice_axis = 0
+    pool.mamba_cache = MambaPool.State(
+        conv=[torch.zeros(NUM_LAYERS, NUM_SLOTS, 4, 5) for _ in range(num_conv)],
+        temporal=temporal,
+    )
+    return pool
+
+
+class TestMambaStateTransferBuffers(unittest.TestCase):
+    def test_conv_only_state_advertises_no_empty_buffer(self):
+        """A ShortConv layer declares a degenerate temporal shape, so the pool
+        allocates an empty tensor for it. The RDMA engine rejects a zero-length
+        region and fails the batch registration that carries the real buffers,
+        so an empty buffer must never be advertised."""
+        pool = _pool(torch.zeros(NUM_LAYERS, NUM_SLOTS, 0, 0, 0))
+
+        _, lens, item_lens = pool.get_contiguous_buf_infos()
+
+        self.assertNotIn(0, lens)
+        self.assertNotIn(0, item_lens)
+        self.assertEqual(len(lens), 2 * NUM_LAYERS)
+
+    def test_temporal_state_is_still_advertised(self):
+        pool = _pool(torch.zeros(NUM_LAYERS, NUM_SLOTS, 6, 7, 8))
+
+        _, lens, _ = pool.get_contiguous_buf_infos()
+
+        self.assertNotIn(0, lens)
+        self.assertEqual(len(lens), 3 * NUM_LAYERS)
+
+    def test_dims_stay_aligned_with_buffers(self):
+        """The per-tensor lists are parallel-indexed, so dropping a buffer has to
+        drop its dim too."""
+        pool = _pool(torch.zeros(NUM_LAYERS, NUM_SLOTS, 0, 0, 0))
+
+        _, lens, _ = pool.get_contiguous_buf_infos()
+
+        self.assertEqual(len(pool.get_state_dim_per_tensor()), len(lens))
+
+
+if __name__ == "__main__":
+    unittest.main()