diff --git a/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py b/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py
index 9d991652a..7bc7cb2b1 100644
--- a/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py
+++ b/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py
@@ -389,16 +389,19 @@ class AscendAttnBackend(AttentionBackend):
         self.graph_mode = False
 
     def init_cuda_graph_state(self, max_bs: int, max_num_tokens: int):
+        total_context_len = self.max_context_len + self.page_size - 1
+        if self.speculative_num_draft_tokens is not None:
+            total_context_len += self.speculative_num_draft_tokens
         self.graph_metadata = {
             "block_tables": torch.empty(
-                (max_bs, (self.max_context_len + self.page_size - 1) // self.page_size),
+                (max_bs, total_context_len // self.page_size),
                 dtype=torch.int32,
                 device=self.device,
             ),
         }
         if self.is_hybrid_swa:
             self.graph_metadata["block_tables_swa"] = torch.empty(
-                (max_bs, (self.max_context_len + self.page_size - 1) // self.page_size),
+                (max_bs, total_context_len // self.page_size),
                 dtype=torch.int32,
                 device=self.device,
             )