Update modeling_qwen3sa.py

modeling_qwen3sa.py

@@ -706,7 +706,7 @@ class Qwen3SummaryAttention(Qwen3Attention):
         else:
             self._sliding_chunk_num = int(val)
 
-        if config.summary_independent_parameters:
+        if config.summary_independent_parameters and config.mix_coeff > 0:
             self.q_proj_summary = nn.Linear(
                 config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
             )