Merge pull request #15190 from AUTOMATIC1111/dora-weight-decompose

Fix built-in lora system bugs caused by torch.nn.MultiheadAttention

extensions-builtin/Lora/network.py

@@ -117,6 +117,12 @@ class NetworkModule:
 
         if hasattr(self.sd_module, 'weight'):
             self.shape = self.sd_module.weight.shape
+        elif isinstance(self.sd_module, nn.MultiheadAttention):
+            # For now, only self-attn use Pytorch's MHA
+            # So assume all qkvo proj have same shape
+            self.shape = self.sd_module.out_proj.weight.shape
+        else:
+            self.shape = None
 
         self.ops = None
         self.extra_kwargs = {}
@@ -146,7 +152,7 @@ class NetworkModule:
         self.alpha = weights.w["alpha"].item() if "alpha" in weights.w else None
         self.scale = weights.w["scale"].item() if "scale" in weights.w else None
 
-        self.dora_scale = weights.w["dora_scale"] if "dora_scale" in weights.w else None
+        self.dora_scale = weights.w.get("dora_scale", None)
         self.dora_mean_dim = tuple(i for i in range(len(self.shape)) if i != 1)
 
     def multiplier(self):

extensions-builtin/Lora/networks.py

@@ -429,9 +429,12 @@ def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn
             if isinstance(self, torch.nn.MultiheadAttention) and module_q and module_k and module_v and module_out:
                 try:
                     with torch.no_grad():
-                        updown_q, _ = module_q.calc_updown(self.in_proj_weight)
-                        updown_k, _ = module_k.calc_updown(self.in_proj_weight)
-                        updown_v, _ = module_v.calc_updown(self.in_proj_weight)
+                        # Send "real" orig_weight into MHA's lora module
+                        qw, kw, vw = self.in_proj_weight.chunk(3, 0)
+                        updown_q, _ = module_q.calc_updown(qw)
+                        updown_k, _ = module_k.calc_updown(kw)
+                        updown_v, _ = module_v.calc_updown(vw)
+                        del qw, kw, vw
                         updown_qkv = torch.vstack([updown_q, updown_k, updown_v])
                         updown_out, ex_bias = module_out.calc_updown(self.out_proj.weight)