Address comment

Signed-off-by: Chang Liu (Enterprise Products) <9713593+chang-l@users.noreply.github.com>
Signed-off-by: Fanrong Li <23290157+lfr-0531@users.noreply.github.com>

Author: chang-l

tensorrt_llm/_torch/attention_backend/sparse/dsa.py

@@ -1114,7 +1114,7 @@ def forward(self, qr: torch.Tensor, hidden_states: torch.Tensor,
                                         weights.to(torch.float32))
 
 
-class DSATrtllmAttention(TrtllmAttention, nn.Module):
+class DSATrtllmAttention(TrtllmAttention):
     Metadata = DSAtrtllmAttentionMetadata
 
     def __init__(
@@ -1151,7 +1151,6 @@ def __init__(
             skip_create_weights_in_init=skip_create_weights_in_init,
             attention_chunk_size=attention_chunk_size,
             **kwargs)
-        nn.Module.__init__(self)
 
         self.indexer = Indexer(quant_config, pos_embd_params, mla_params,
                                skip_create_weights_in_init,

tensorrt_llm/_torch/models/modeling_deepseekv3.py

@@ -383,8 +383,6 @@ def split_kv_b_proj(kv_b_proj: torch.Tensor,
                     continue
                 elif names[-1] == "next_layer_layernorm":
                     continue
-                elif names[-1] == "indexer" or 'mqa.indexer' in name:
-                    continue
                 else:
                     module_weights = filter_weights(name, weights)
                     if hasattr(module, 'load_weights'):