diff --git a/python/sglang/srt/layers/attention/trtllm_mha_backend.py b/python/sglang/srt/layers/attention/trtllm_mha_backend.py
index e270f2ccb..068791b02 100644
--- a/python/sglang/srt/layers/attention/trtllm_mha_backend.py
+++ b/python/sglang/srt/layers/attention/trtllm_mha_backend.py
@@ -324,6 +324,9 @@ class TRTLLMHAAttnBackend(FlashInferAttnBackend):
                 metadata.cache_seqlens_int32 = self.decode_cuda_graph_metadata[
                     "cache_seqlens"
                 ][:bs]
+                metadata.cache_seqlens_int32.copy_(
+                    seq_lens + self.speculative_step_id + 1
+                )
                 metadata.max_seq_len_k = seq_lens.max().item() + (
                     self.speculative_step_id + 1
                 )