test2_强制修改为ALL

vllm/model_executor/models/adapters_custom/adapters_classify.py

@@ -322,7 +322,7 @@ def new_hy_05b_dense_official_classification(cls: _T) -> _T:
                     quant_config=quant_config,
                     params_dtype=torch.float32,
                     prefix=maybe_prefix(prefix, "pool_head2"),
-                    return_bias=True,
+                    return_bias=False,
                 )
                 # 兼容 ForSequenceClassification：将 score 直接指向最终分类头
                 # 不再单独创建一层；pool_head2 即最终打分层
@@ -457,7 +457,7 @@ def hy_2b_dense_classification_official_hf_multihead_full_mask(cls: _T) -> _T:
                     quant_config=quant_config,
                     params_dtype=torch.float32,
                     prefix=maybe_prefix(prefix, "pool_head2"),
-                    return_bias=True,
+                    return_bias=False,
                 )
                 self.qfeat_emb =ReplicatedLinear(
                     2,
@@ -514,9 +514,10 @@ def hy_2b_dense_classification_official_hf_multihead_full_mask(cls: _T) -> _T:
                     "PoolerConfig must be provided to use classification head")
 
                 # Determine pooling type (fallback to config.pool_type)
-                pooling_type_str = (pooler_config.pooling_type
-                                    if pooler_config.pooling_type is not None
-                                    else getattr(config, "pool_type", "LAST")).upper()
+                # pooling_type_str = (pooler_config.pooling_type
+                #                     if pooler_config.pooling_type is not None
+                #                     else getattr(config, "pool_type", "LAST")).upper()
+                pooling_type_str="ALL"
                 if pooling_type_str == "LASTTOKEN":
                     pooling_type_str = "LAST"
                 pooling_type = PoolingType[pooling_type_str]
@@ -552,10 +553,10 @@ def hy_2b_dense_classification_official_hf_multihead_full_mask(cls: _T) -> _T:
             a_wei = self.qfeat_fc2(qhidden)
             a_bias = self.qfeat_fc3(qhidden)
 
-            sat_logits = pooled_output_sat[:,-1]
-            auth_logits = pooled_output_auth[:,-2]
-            time_logits = pooled_output_time[:,-3]
-            rel_logits = pooled_output_rel[:,-4]
+            sat_logits = pooled_output_sat[:,-1,:]
+            auth_logits = pooled_output_auth[:,-2,:]
+            time_logits = pooled_output_time[:,-3,:]
+            rel_logits = pooled_output_rel[:,-4,:]
             multii_logits = torch.concat([rel_logits, time_logits, auth_logits], dim=1)
             task_logits = (a_wei * multii_logits + a_bias).sum(dim=1, keepdim=True)
             task_logits = torch.sigmoid(task_logits)
@@ -586,7 +587,7 @@ def hy_2b_dense_classification_official_hf_multihead_full_mask(cls: _T) -> _T:
 
 
 
-            return  pooled_output#reward
+            return  reward
 
         def forward(
             self,

vllm/version.py

@@ -2,10 +2,10 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 try:
-    # __version__ = "0.11.0"
-    # __version_tuple__ = (0, 11, 0)
-    # __hcu_version__ = f'0.11.0+das.opt1.alpha.c16e075.dtk25042'
-    # from vllm.version import __version__, __version_tuple__, __hcu_version__
+    #__version__ = "0.11.0"
+    #__version_tuple__ = (0, 11, 0)
+    #__hcu_version__ = f'0.11.0+das.opt1.alpha.c16e075.dtk25042'
+    #from vllm.version import __version__, __version_tuple__, __hcu_version__
     from ._version import __version__, __version_tuple__
 except Exception as e:
     import warnings