diff --git a/python/sglang/srt/layers/attention/trtllm_mha_backend.py b/python/sglang/srt/layers/attention/trtllm_mha_backend.py
index d58f3f661..006911a48 100644
--- a/python/sglang/srt/layers/attention/trtllm_mha_backend.py
+++ b/python/sglang/srt/layers/attention/trtllm_mha_backend.py
@@ -714,12 +714,7 @@ class TRTLLMHAAttnBackend(FlashInferAttnBackend):
             )
             # Query-side max length, sourced from the host-resident extend lengths
             # (sync-free); for plain prefill these equal the full seq lens.
-            # NOTE: in piecewise CUDA graph warmup, extend_seq_lens_cpu is a torch.Tensor;
-            # Python's max() returns a 0-d tensor, but flashinfer expects an int.
-            max_q = max(forward_batch.extend_seq_lens_cpu)
-            metadata.max_seq_len_q = (
-                int(max_q.item()) if isinstance(max_q, torch.Tensor) else int(max_q)
-            )
+            metadata.max_seq_len_q = int(max(forward_batch.extend_seq_lens_cpu))
             if (
                 forward_batch.extend_prefix_lens_cpu is not None
                 and any(forward_batch.extend_prefix_lens_cpu)
diff --git a/python/sglang/srt/model_executor/forward_batch_info.py b/python/sglang/srt/model_executor/forward_batch_info.py
index 8a3a7b02c..1df55a02f 100644
--- a/python/sglang/srt/model_executor/forward_batch_info.py
+++ b/python/sglang/srt/model_executor/forward_batch_info.py
@@ -1267,8 +1267,8 @@ class ForwardBatch(ForwardBatchDeepSeekMHAMixin):
                     self.extend_start_loc = torch.arange(
                         bs, dtype=torch.int32, device=self.seq_lens.device
                     )
-                    self.extend_prefix_lens_cpu = self.extend_prefix_lens.cpu()
-                    self.extend_seq_lens_cpu = self.extend_seq_lens.cpu()
+                    self.extend_prefix_lens_cpu = self.extend_prefix_lens.cpu().tolist()
+                    self.extend_seq_lens_cpu = self.extend_seq_lens.cpu().tolist()
                     self.extend_logprob_start_lens_cpu = self.extend_prefix_lens_cpu
             else:
                 if self.spec_info is not None:
diff --git a/python/sglang/srt/model_executor/runner/prefill_cuda_graph_runner.py b/python/sglang/srt/model_executor/runner/prefill_cuda_graph_runner.py
index c6834a951..94970e151 100644
--- a/python/sglang/srt/model_executor/runner/prefill_cuda_graph_runner.py
+++ b/python/sglang/srt/model_executor/runner/prefill_cuda_graph_runner.py
@@ -728,11 +728,9 @@ class PrefillCudaGraphRunner(BaseCudaGraphRunner):
                 extend_seq_lens=shape_inputs["extend_seq_lens"],
                 extend_prefix_lens=shape_inputs["extend_prefix_lens"],
                 extend_start_loc=shape_inputs["extend_start_loc"],
-                extend_prefix_lens_cpu=torch.zeros(
-                    (bs,), dtype=torch.int64, device="cpu"
-                ),
-                extend_seq_lens_cpu=torch.tensor(lens_cpu, device="cpu"),
-                extend_logprob_start_lens_cpu=torch.tensor(lens_cpu, device="cpu"),
+                extend_prefix_lens_cpu=[0] * bs,
+                extend_seq_lens_cpu=list(lens_cpu),
+                extend_logprob_start_lens_cpu=list(lens_cpu),
                 positions=_slot("positions"),
                 global_num_tokens_gpu=global_num_tokens_gpu,
                 global_num_tokens_for_logprob_gpu=global_num_tokens_for_logprob_gpu,
diff --git a/test/manual/attention/test_flashattn_backend.py b/test/manual/attention/test_flashattn_backend.py
index 5b99abceb..51c04968d 100644
--- a/test/manual/attention/test_flashattn_backend.py
+++ b/test/manual/attention/test_flashattn_backend.py
@@ -222,15 +222,11 @@ class TestFlashAttentionBackend(CustomTestCase):
                 extend_prefix_lens=torch.tensor(
                     [prefix_len] * self.batch_size, device=self.device
                 ),
-                extend_prefix_lens_cpu=torch.tensor(
-                    [prefix_len] * self.batch_size, device="cpu"
-                ),
+                extend_prefix_lens_cpu=[prefix_len] * self.batch_size,
                 extend_seq_lens=torch.tensor(
                     [q_len] * self.batch_size, device=self.device
                 ),
-                extend_seq_lens_cpu=torch.tensor(
-                    [q_len] * self.batch_size, device="cpu"
-                ),
+                extend_seq_lens_cpu=[q_len] * self.batch_size,
             )
             if attn_cp_size > 1:
                 forward_batch.attn_cp_metadata = type(
diff --git a/test/manual/attention/test_flashattn_mla_backend.py b/test/manual/attention/test_flashattn_mla_backend.py
index fb44bf0f3..d19f992b6 100644
--- a/test/manual/attention/test_flashattn_mla_backend.py
+++ b/test/manual/attention/test_flashattn_mla_backend.py
@@ -190,15 +190,11 @@ class TestFlashAttentionMLABackend(CustomTestCase):
                 extend_prefix_lens=torch.tensor(
                     [prefix_len] * self.batch_size, device=self.device
                 ),
-                extend_prefix_lens_cpu=torch.tensor(
-                    [prefix_len] * self.batch_size, device="cpu"
-                ),
+                extend_prefix_lens_cpu=[prefix_len] * self.batch_size,
                 extend_seq_lens=torch.tensor(
                     [q_len] * self.batch_size, device=self.device
                 ),
-                extend_seq_lens_cpu=torch.tensor(
-                    [q_len] * self.batch_size, device="cpu"
-                ),
+                extend_seq_lens_cpu=[q_len] * self.batch_size,
             )
 
         else:  # ForwardMode.DECODE
diff --git a/test/registered/kernels/test_dsa_indexer.py b/test/registered/kernels/test_dsa_indexer.py
index 90d4bd9c9..3102f2126 100644
--- a/test/registered/kernels/test_dsa_indexer.py
+++ b/test/registered/kernels/test_dsa_indexer.py
@@ -367,11 +367,9 @@ class TestDSAIndexer(CustomTestCase):
                 extend_prefix_lens=torch.tensor(
                     [total_len - q_len] * batch_size, device=self.device
                 ),
-                extend_prefix_lens_cpu=torch.tensor(
-                    [total_len - q_len] * batch_size, device="cpu"
-                ),
+                extend_prefix_lens_cpu=[total_len - q_len] * batch_size,
                 extend_seq_lens=torch.tensor([q_len] * batch_size, device=self.device),
-                extend_seq_lens_cpu=torch.tensor([q_len] * batch_size, device="cpu"),
+                extend_seq_lens_cpu=[q_len] * batch_size,
             )
         else:  # ForwardMode.DECODE
             decode_len = 1