diff --git a/vllm/model_executor/layers/mamba/gdn/qwen_gdn_linear_attn.py b/vllm/model_executor/layers/mamba/gdn/qwen_gdn_linear_attn.py index 60879f8fb1..f3ebcd6325 100644 --- a/vllm/model_executor/layers/mamba/gdn/qwen_gdn_linear_attn.py +++ b/vllm/model_executor/layers/mamba/gdn/qwen_gdn_linear_attn.py @@ -1252,9 +1252,13 @@ class QwenGatedDeltaNetAttention(GatedDeltaNetAttention): if spec_sequence_masks is not None: if attn_metadata.num_prefills == 0 and attn_metadata.num_decodes == 0: mixed_qkv_spec = mixed_qkv + a_spec = a + b_spec = b mixed_qkv_non_spec = None else: mixed_qkv_spec = mixed_qkv.index_select(0, spec_token_indx) + a_spec = a.index_select(0, spec_token_indx) + b_spec = b.index_select(0, spec_token_indx) mixed_qkv_non_spec = mixed_qkv.index_select(0, non_spec_token_indx) else: mixed_qkv_spec = None @@ -1379,8 +1383,8 @@ class QwenGatedDeltaNetAttention(GatedDeltaNetAttention): core_attn_out_spec, last_recurrent_state = ( fused_sigmoid_gating_delta_rule_update( A_log=self.A_log, - a=a, - b=b, + a=a_spec, + b=b_spec, dt_bias=self.dt_bias, q=query_spec, k=key_spec,