diff --git a/python/sglang/srt/layers/attention/trtllm_mha_backend.py b/python/sglang/srt/layers/attention/trtllm_mha_backend.py
index 17f565ef1..93a028d20 100644
--- a/python/sglang/srt/layers/attention/trtllm_mha_backend.py
+++ b/python/sglang/srt/layers/attention/trtllm_mha_backend.py
@@ -484,6 +484,7 @@ class TRTLLMHAAttnBackend(FlashInferAttnBackend):
             ]
             metadata.page_table[:, :max_seq_pages].copy_(page_indices // self.page_size)
             self._copy_swa_page_table(metadata, page_indices, max_seq_pages)
+            metadata.max_seq_len_q = self.speculative_num_draft_tokens
         elif forward_mode.is_draft_extend(include_v2=True):
             metadata = self.draft_extend_metadata[bs]
             metadata.cache_seqlens_int32.copy_(seq_lens)