diff --git a/python/sglang/srt/layers/attention/dsa/dsa_indexer.py b/python/sglang/srt/layers/attention/dsa/dsa_indexer.py
index 1fb37764c..90cf90637 100644
--- a/python/sglang/srt/layers/attention/dsa/dsa_indexer.py
+++ b/python/sglang/srt/layers/attention/dsa/dsa_indexer.py
@@ -1835,8 +1835,7 @@ class Indexer(MultiPlatformOp):
             # creates a Dynamo shape guard. These graph modes never have empty
             # batches.
             if not in_piecewise_or_breakable_cuda_graph:
-                assert forward_batch.seq_lens_cpu is not None
-                if len(forward_batch.seq_lens_cpu) == 0:
+                if forward_batch.seq_lens.numel() == 0:
                     # this seems b/c max-pad, no worries?
                     # if x.shape[0] != 0:
                     #     print(
diff --git a/python/sglang/srt/layers/attention/dsa_backend.py b/python/sglang/srt/layers/attention/dsa_backend.py
index 341c3a2dc..1b8f70e7e 100644
--- a/python/sglang/srt/layers/attention/dsa_backend.py
+++ b/python/sglang/srt/layers/attention/dsa_backend.py
@@ -755,6 +755,14 @@ class DeepseekSparseAttnBackend(
                 page_table, repeats=self.speculative_num_draft_tokens, dim=0
             )
         elif forward_batch.forward_mode.is_draft_extend_v2():
+            if forward_batch.extend_prefix_lens_cpu is None:
+                assert forward_batch.extend_prefix_lens is not None
+                forward_batch.extend_prefix_lens_cpu = (
+                    forward_batch.extend_prefix_lens.cpu().tolist()
+                )
+            if forward_batch.seq_lens_cpu is None:
+                forward_batch.seq_lens_cpu = forward_batch.seq_lens.cpu()
+                forward_batch.seq_lens_sum = int(forward_batch.seq_lens_cpu.sum())
             assert (
                 forward_batch.extend_seq_lens_cpu is not None
                 and forward_batch.extend_seq_lens is not None