Update modeling_phi3.py

modeling_phi3.py

@@ -328,9 +328,9 @@ class Phi3Attention(nn.Module):
         self.attncut = True
         self.headcut = True
         self.layercut = False
-        self.layercut_idx = 24
-        self.offset = 70
-        head_num=24
+        self.layercut_idx = 24 # number of layer kept  
+        self.offset = 70 # system prompt token length
+        head_num=24 # num of heads kept on average
         self.mask = torch.load("headcut_mask/internvl2.0_4B/mask_"+str(head_num)+".pth")
         
     def _init_rope(self):