fix: keep CP HiCache load host indices on CPU

python/sglang/srt/managers/cache_controller.py

@@ -809,7 +809,7 @@ class HiCacheController:
             physical_chunks.append(
                 self.cp_shared_kv_layout.logical_locs_to_physical(selected_logical_locs)
             )
-            host_chunks.append(node.cp_hicache.host_indices.to(device_indices.device))
+            host_chunks.append(node.cp_hicache.host_indices)
 
         if not host_chunks:
             self._append_completed_load_ack(node_id)

test/registered/unit/managers/test_hicache_controller_cp.py

@@ -93,6 +93,15 @@ class FakeAllocator:
         return self.alloc_result[:need_size].clone()
 
 
+class HostIndicesTensor(torch.Tensor):
+    @staticmethod
+    def __new__(cls, data):
+        return torch.Tensor._make_subclass(cls, data, require_grad=False)
+
+    def to(self, *args, **kwargs):
+        raise AssertionError("load_cp should not move host indices before queuing")
+
+
 class DummyEvent:
     def record(self):
         pass
@@ -246,6 +255,29 @@ class TestHiCacheControllerCPLoad(TestHiCacheControllerCPWrite):
         self.assertEqual(host_pool.loads, [])
         self.assertEqual(len(controller.ack_load_queue), 1)
 
+    def test_cp_load_queues_cpu_host_indices_before_backend_moves(self):
+        host_pool = FakeHostPool(torch.tensor([100, 101, 102, 103], dtype=torch.int64))
+        allocator = FakeAllocator(alloc_result=torch.arange(64, 80, dtype=torch.int64))
+        controller = self.make_controller(host_pool, allocator=allocator, cp_rank=1)
+        host_indices = HostIndicesTensor(torch.tensor([100, 101, 102, 103], dtype=torch.int64))
+        node = TreeNode()
+        node.host_len = 16
+        node.cp_hicache = type(
+            "CpHiCacheMetadataStub",
+            (),
+            {
+                "owned_positions": torch.tensor([4, 5, 6, 7], dtype=torch.int64),
+                "host_indices": host_indices,
+            },
+        )()
+
+        controller.load_cp([node], node_id=13)
+
+        queued_op = controller.load_queue[0]
+        self.assertEqual(queued_op.host_indices.device.type, "cpu")
+        self.assertEqual(queued_op.host_indices.tolist(), [100, 101, 102, 103])
+        self.assertEqual(queued_op.device_indices.tolist(), [20, 21, 22, 23])
+
 
 if __name__ == "__main__":
     main()