[HiCache] fix: Mooncake Dummy Client mode for hybrid Mamba models (#25278)

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Teng Ma <stmatengss@users.noreply.github.com>

python/sglang/srt/mem_cache/storage/mooncake_store/mooncake_store.py

@@ -296,6 +296,52 @@ class MooncakeBaseStore:
 
 class MooncakeStore(HiCacheStorage, MooncakeBaseStore):
 
+    @staticmethod
+    def _standalone_required_bytes(mem_pool: Any) -> int:
+        """Compute total bytes of host buffers that must be visible to the real client.
+
+        In standalone (dummy client) mode, the real mooncake_client process needs
+        to map any host buffers we will later pass by pointer via register_buffer().
+        For hybrid models, that includes KV + sidecar pools (e.g. Mamba temporal/conv).
+        """
+        # Prefer a generic "hybrid pool" accessor when present.
+        total = 0
+        seen_ptrs: set[int] = set()
+
+        def _add_tensor(t: Optional[torch.Tensor]):
+            nonlocal total
+            if t is None:
+                return
+            try:
+                ptr = int(t.data_ptr())
+            except Exception:
+                return
+            if ptr in seen_ptrs:
+                return
+            seen_ptrs.add(ptr)
+            total += int(t.numel() * t.element_size())
+
+        # Always include the anchor KV buffer if present.
+        _add_tensor(getattr(mem_pool, "kv_buffer", None))
+
+        # HostPoolGroup: include each pool's hybrid buffers when available.
+        entries = getattr(mem_pool, "entries", None)
+        if entries:
+            for entry in entries:
+                host_pool = getattr(entry, "host_pool", None)
+                if host_pool is None:
+                    continue
+                # KV pool anchor memory is already covered, but harmless if added twice.
+                _add_tensor(getattr(host_pool, "kv_buffer", None))
+                for buf in getattr(host_pool, "get_hybrid_pool_buffer", lambda: [])():
+                    _add_tensor(buf)
+            return total
+
+        # Single HostKVCache-like pool: add its sidecar buffers if any.
+        for buf in getattr(mem_pool, "get_hybrid_pool_buffer", lambda: [])():
+            _add_tensor(buf)
+        return total
+
     def __init__(
         self, storage_config: HiCacheStorageConfig = None, mem_pool: HostKVCache = None
     ):
@@ -351,8 +397,9 @@ class MooncakeStore(HiCacheStorage, MooncakeBaseStore):
                         "Please set standalone_storage=False "
                         "or upgrade Mooncake by 'pip install mooncake --upgrade'."
                     )
+                required_bytes = self._standalone_required_bytes(mem_pool)
                 ret_code = self.store.setup_dummy(
-                    mem_pool.size * mem_pool.size_per_token,
+                    required_bytes,
                     DEFAULT_LOCAL_BUFFER_SIZE,  # Zero copy interface does not need local buffer
                     self.config.client_server_address,
                 )