fix set quantization

Author: shihaobai

lightllm/common/basemodel/layer_weights/meta_weights/__init__.py

@@ -1,5 +1,6 @@
 from .base_weight import BaseWeight
 from .mm_weight import (
+    MMWeightTpl,
     MMWeight,
     MultiMMWeight,
     ROWMMWeight,

lightllm/common/basemodel/layer_weights/meta_weights/mm_weight.py

@@ -30,6 +30,7 @@ def mm(self, input_tensor, out=None, use_custom_tensor_mananger=True):
         return torch.addmm(self.bias, input_tensor, self.weight, out=out)
 
     def _post_load_weights(self):
+        print(self.quant_method)
         if self.quant_method is not None:
             self.weight = self.quant_method.quantize(self.weight.cuda(self.tp_rank_))
             return

lightllm/common/basemodel/layer_weights/transformer_layer_weight.py

@@ -2,7 +2,7 @@
 
 # from lightllm.common.layers.mm import MM
 from .base_layer_weight import BaseLayerWeight
-from .meta_weights import BaseWeight, MultiMMWeight, MMWeight, FusedMoeWeight
+from .meta_weights import BaseWeight, MultiMMWeight, MMWeightTpl, FusedMoeWeight
 from lightllm.utils.log_utils import init_logger
 
 logger = init_logger(__name__)
@@ -51,7 +51,7 @@ def set_quantization(self):
         mix_quant_list = self.quant_cfg.get_mixed_list(self.layer_num_)
         for attr_name in dir(self):
             attr = getattr(self, attr_name)
-            if isinstance(attr, MMWeight) or isinstance(attr, FusedMoeWeight):
+            if isinstance(attr, MMWeightTpl) or isinstance(attr, FusedMoeWeight):
                 if attr_name in mix_quant_list:
                     attr.set_quant_method(self.quant_cfg.get_quant_method(self.layer_num_, attr_name))
                     attr_quant_type = self.quant_cfg.get_quant_type(self.layer_num_, attr_name)