diff --git a/python/sglang/srt/models/deepseek_v2.py b/python/sglang/srt/models/deepseek_v2.py
index b299c3037..241c46cf9 100644
--- a/python/sglang/srt/models/deepseek_v2.py
+++ b/python/sglang/srt/models/deepseek_v2.py
@@ -583,13 +583,17 @@ class DeepseekV2AttentionMLA(nn.Module):
                 return AttnForwardMethod.MLA
         elif self.attention_backend == "fa3":
             # Flash Attention: Use MHA with chunked KV cache when prefilling on long sequences.
+            if forward_batch.extend_prefix_lens_cpu is not None:
+                sum_extend_prefix_lens = sum(forward_batch.extend_prefix_lens_cpu)
             if (
                 forward_batch.forward_mode.is_extend()
                 and not self.disable_chunked_prefix_cache
                 and not forward_batch.forward_mode.is_target_verify()
                 and not forward_batch.forward_mode.is_draft_extend()
-                and sum(forward_batch.extend_prefix_lens_cpu)
-                >= self.chunked_prefix_cache_threshold
+                and (
+                    sum_extend_prefix_lens >= self.chunked_prefix_cache_threshold
+                    or sum_extend_prefix_lens == 0
+                )
             ):
                 return AttnForwardMethod.MHA_CHUNKED_KV
             else: