Merge branch 'dev' into torch210

2 files changed, 13 insertions, 27 deletions

diff --git a/extensions-builtin/Lora/network_oft.py b/extensions-builtin/Lora/network_oft.py
index 05c37811..fa647020 100644
--- a/extensions-builtin/Lora/network_oft.py
+++ b/extensions-builtin/Lora/network_oft.py
@@ -21,6 +21,8 @@ class NetworkModuleOFT(network.NetworkModule):
         self.lin_module = None
         self.org_module: list[torch.Module] = [self.sd_module]
 
+        self.scale = 1.0
+
         # kohya-ss
         if "oft_blocks" in weights.w.keys():
             self.is_kohya = True
@@ -53,12 +55,18 @@ class NetworkModuleOFT(network.NetworkModule):
             self.constraint = None
             self.block_size, self.num_blocks = factorization(self.out_dim, self.dim)
 
-    def calc_updown_kb(self, orig_weight, multiplier):
+    def calc_updown(self, orig_weight):
         oft_blocks = self.oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype)
-        oft_blocks = oft_blocks - oft_blocks.transpose(1, 2) # ensure skew-symmetric orthogonal matrix
+        eye = torch.eye(self.block_size, device=self.oft_blocks.device)
+
+        if self.is_kohya:
+            block_Q = oft_blocks - oft_blocks.transpose(1, 2) # ensure skew-symmetric orthogonal matrix
+            norm_Q = torch.norm(block_Q.flatten())
+            new_norm_Q = torch.clamp(norm_Q, max=self.constraint)
+            block_Q = block_Q * ((new_norm_Q + 1e-8) / (norm_Q + 1e-8))
+            oft_blocks = torch.matmul(eye + block_Q, (eye - block_Q).float().inverse())
 
         R = oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype)
-        R = R * multiplier + torch.eye(self.block_size, device=orig_weight.device)
 
         # This errors out for MultiheadAttention, might need to be handled up-stream
         merged_weight = rearrange(orig_weight, '(k n) ... -> k n ...', k=self.num_blocks, n=self.block_size)
@@ -72,26 +80,3 @@ class NetworkModuleOFT(network.NetworkModule):
         updown = merged_weight.to(orig_weight.device, dtype=orig_weight.dtype) - orig_weight
         output_shape = orig_weight.shape
         return self.finalize_updown(updown, orig_weight, output_shape)
-
-    def calc_updown(self, orig_weight):
-        # if alpha is a very small number as in coft, calc_scale() will return a almost zero number so we ignore it
-        multiplier = self.multiplier()
-        return self.calc_updown_kb(orig_weight, multiplier)
-
-    # override to remove the multiplier/scale factor; it's already multiplied in get_weight
-    def finalize_updown(self, updown, orig_weight, output_shape, ex_bias=None):
-        if self.bias is not None:
-            updown = updown.reshape(self.bias.shape)
-            updown += self.bias.to(orig_weight.device, dtype=orig_weight.dtype)
-            updown = updown.reshape(output_shape)
-
-        if len(output_shape) == 4:
-            updown = updown.reshape(output_shape)
-
-        if orig_weight.size().numel() == updown.size().numel():
-            updown = updown.reshape(orig_weight.shape)
-
-        if ex_bias is not None:
-            ex_bias = ex_bias * self.multiplier()
-
-        return updown, ex_bias
diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py
index 7f814706..629bf853 100644
--- a/extensions-builtin/Lora/networks.py
+++ b/extensions-builtin/Lora/networks.py
@@ -159,7 +159,8 @@ def load_network(name, network_on_disk):
     bundle_embeddings = {}

 

     for key_network, weight in sd.items():

-        key_network_without_network_parts, network_part = key_network.split(".", 1)

+        key_network_without_network_parts, _, network_part = key_network.partition(".")

+

         if key_network_without_network_parts == "bundle_emb":

             emb_name, vec_name = network_part.split(".", 1)

             emb_dict = bundle_embeddings.get(emb_name, {})