diff --git a/python/sglang/srt/layers/attention/minicpm_backend.py b/python/sglang/srt/layers/attention/minicpm_backend.py
index 70683277c0ce..473dd6841010 100644
--- a/python/sglang/srt/layers/attention/minicpm_backend.py
+++ b/python/sglang/srt/layers/attention/minicpm_backend.py
@@ -1756,6 +1756,19 @@ def init_forward_metadata_replay_cuda_graph(
                 kv_indptr_view = self.decode_cuda_graph_metadata[
                     "flashinfer_kv_indptr"
                 ][: sparse_bs + 1]
+                kv_indptr_view[0] = 0
+                if sparse_real_bs > 0:
+                    actual_seqlens = metadata.sparse_cache_seqlens_int32[
+                        :sparse_real_bs
+                    ].clone()
+                    actual_seqlens = torch.clamp(
+                        actual_seqlens, max=self.num_sparse_topk_tokens
+                    )
+                    kv_indptr_view[1 : sparse_real_bs + 1] = torch.cumsum(
+                        actual_seqlens, dim=0
+                    )
+                kv_indptr_view[sparse_real_bs:].fill_(kv_indptr_view[sparse_real_bs])
+
                 # kv_indices only needs num_sparse_topk_tokens per batch
                 kv_indices_view = self.decode_cuda_graph_metadata[
                     "flashinfer_kv_indices"
@@ -1763,7 +1776,6 @@ def init_forward_metadata_replay_cuda_graph(
                 kv_last_page_len_view = self.decode_cuda_graph_metadata[
                     "flashinfer_kv_last_page_len"
                 ][:sparse_bs]
-                kv_indptr_view[sparse_real_bs:].fill_(kv_indptr_view[-1])
                 kv_last_page_len_view[sparse_real_bs:].fill_(0)
 
                 # Retrieve the wrapper stored during capture