diff --git a/python/sglang/srt/models/lfm2.py b/python/sglang/srt/models/lfm2.py
index 3694e32fa..2750a0f81 100644
--- a/python/sglang/srt/models/lfm2.py
+++ b/python/sglang/srt/models/lfm2.py
@@ -302,11 +302,13 @@ class Lfm2ShortConv(nn.Module):
                     ]
                 )
                 cache_indices = mamba_indices.to(torch.int32)
+                has_initial_state = forward_batch.extend_prefix_lens > 0
             else:
                 query_start_loc = torch.tensor(
                     [0, T], dtype=torch.int32, device=hidden_states.device
                 )
                 cache_indices = mamba_indices[:1].to(torch.int32)
+                has_initial_state = forward_batch.extend_prefix_lens[:1] > 0
 
             conv_out = causal_conv1d_fn(
                 Bx_t,
@@ -314,7 +316,7 @@ class Lfm2ShortConv(nn.Module):
                 self.conv_bias,
                 query_start_loc=query_start_loc,
                 cache_indices=cache_indices,
-                has_initial_state=None,
+                has_initial_state=has_initial_state,
                 conv_states=conv_state,
                 activation=None,
             ).transpose(0, 1)
diff --git a/python/sglang/srt/models/lfm2_moe.py b/python/sglang/srt/models/lfm2_moe.py
index 4c7d5d06d..fcc396357 100644
--- a/python/sglang/srt/models/lfm2_moe.py
+++ b/python/sglang/srt/models/lfm2_moe.py
@@ -360,10 +360,12 @@ class Lfm2MoeShortConv(nn.Module):
                 query_start_loc[:-1] = extend_start_loc
                 query_start_loc[-1] = T
                 cache_indices = mamba_indices.to(torch.int32)
+                has_initial_state = forward_batch.extend_prefix_lens > 0
             else:
                 # Single sequence: [0, T]
                 query_start_loc = hidden_states.new_tensor([0, T], dtype=torch.int32)
                 cache_indices = mamba_indices[:1].to(torch.int32)
+                has_initial_state = forward_batch.extend_prefix_lens[:1] > 0
 
             conv_out = causal_conv1d_fn(
                 Bx_t,
@@ -371,7 +373,7 @@ class Lfm2MoeShortConv(nn.Module):
                 self.conv_bias,
                 query_start_loc=query_start_loc,
                 cache_indices=cache_indices,
-                has_initial_state=None,
+                has_initial_state=has_initial_state,
                 conv_states=conv_state,
                 activation=None,
             ).transpose(0, 1)