Fix the bug in sfa-cp when MTP is enabled.

Signed-off-by: zzhx1 <[email protected]>

Author: zzhx1

vllm_ascend/attention/sfa_v1.py

@@ -355,16 +355,18 @@ def __init__(
             self._replace_linear_class_for_sfa_cp()
             from vllm_ascend.distributed.parallel_state import \
                 get_shared_weight_group
-            register_layer_to_shared_weight_series(
-                series_name="q_proj",
-                group=get_shared_weight_group(),
-                layer=self.q_proj,
-                prefetch_step=1)
-            register_layer_to_shared_weight_series(
-                series_name="o_proj",
-                group=get_shared_weight_group(),
-                layer=self.o_proj,
-                prefetch_step=1)
+            if is_hidden_layer(self.model_config.hf_config, self.q_proj):
+                register_layer_to_shared_weight_series(
+                    series_name="q_proj",
+                    group=get_shared_weight_group(),
+                    layer=self.q_proj,
+                    prefetch_step=1)
+            if is_hidden_layer(self.model_config.hf_config, self.o_proj):
+                register_layer_to_shared_weight_series(
+                    series_name="o_proj",
+                    group=get_shared_weight_group(),
+                    layer=self.o_proj,
+                    prefetch_step=1)
 
         # indexer param
         self.n_head: int = self.indexer.n_head  # 64