diff --git a/python/sglang/srt/models/qwen3.py b/python/sglang/srt/models/qwen3.py
index 95aaaa235..b6b955ae6 100644
--- a/python/sglang/srt/models/qwen3.py
+++ b/python/sglang/srt/models/qwen3.py
@@ -181,7 +181,10 @@ class Qwen3Attention(nn.Module):
         if get_global_server_args().rl_on_policy_target is not None:
             hidden_states = hidden_states.bfloat16()
 
-        if not _is_npu or forward_batch.forward_mode.is_extend():
+        if (
+            not _is_npu
+            or forward_batch.forward_mode.is_extend_or_draft_extend_or_mixed()
+        ):
             q, k, v = self.forward_prepare_native(
                 positions=positions,
                 hidden_states=hidden_states,
diff --git a/python/sglang/srt/models/qwen3_moe.py b/python/sglang/srt/models/qwen3_moe.py
index 3fcf0cfa0..845502fe2 100644
--- a/python/sglang/srt/models/qwen3_moe.py
+++ b/python/sglang/srt/models/qwen3_moe.py
@@ -620,7 +620,10 @@ class Qwen3MoeAttention(nn.Module):
     ):
         if hidden_states.shape[0] == 0:
             return hidden_states, forward_batch, None
-        if not _is_npu or forward_batch.forward_mode.is_extend():
+        if (
+            not _is_npu
+            or forward_batch.forward_mode.is_extend_or_draft_extend_or_mixed()
+        ):
             return self.forward_prepare_native(
                 positions=positions,
                 hidden_states=hidden_states,