inference working but SLOW

extensions-builtin/Lora/network_oft.py

@@ -12,6 +12,7 @@ class ModuleTypeOFT(network.ModuleType):
 # adapted from https://github.com/kohya-ss/sd-scripts/blob/main/networks/oft.py
 class NetworkModuleOFT(network.NetworkModule):
     def __init__(self,  net: network.Network, weights: network.NetworkWeights):
+
         super().__init__(net, weights)
 
         self.oft_blocks = weights.w["oft_blocks"]
@@ -20,24 +21,29 @@ class NetworkModuleOFT(network.NetworkModule):
         self.dim = self.oft_blocks.shape[0]
         self.num_blocks = self.dim
 
-        #if type(self.alpha) == torch.Tensor:
-        #    self.alpha = self.alpha.detach().numpy()
-
         if "Linear" in self.sd_module.__class__.__name__:
             self.out_dim = self.sd_module.out_features
         elif "Conv" in self.sd_module.__class__.__name__:
             self.out_dim = self.sd_module.out_channels
 
-        self.constraint = self.alpha * self.out_dim
+        self.constraint = self.alpha
+        #self.constraint = self.alpha * self.out_dim
         self.block_size = self.out_dim // self.num_blocks
 
-        self.oft_multiplier = self.multiplier()
+        self.org_module: list[torch.Module] = [self.sd_module]
 
-        # replace forward method of original linear rather than replacing the module
+        self.R = self.get_weight()
-        # self.org_forward = self.sd_module.forward
+
-        # self.sd_module.forward = self.forward
+        self.apply_to()
+
+    # replace forward method of original linear rather than replacing the module
+    def apply_to(self):
+        self.org_forward = self.org_module[0].forward
+        self.org_module[0].forward = self.forward
     
-    def get_weight(self):
+    def get_weight(self, multiplier=None):
+        if not multiplier:
+            multiplier = self.multiplier()
         block_Q = self.oft_blocks - self.oft_blocks.transpose(1, 2)
         norm_Q = torch.norm(block_Q.flatten())
         new_norm_Q = torch.clamp(norm_Q, max=self.constraint)
@@ -45,38 +51,31 @@ class NetworkModuleOFT(network.NetworkModule):
         I = torch.eye(self.block_size, device=self.oft_blocks.device).unsqueeze(0).repeat(self.num_blocks, 1, 1)
         block_R = torch.matmul(I + block_Q, (I - block_Q).inverse())
 
-        block_R_weighted = self.oft_multiplier * block_R + (1 - self.oft_multiplier) * I
+        block_R_weighted = multiplier * block_R + (1 - multiplier) * I
         R = torch.block_diag(*block_R_weighted)
 
         return R
 
     def calc_updown(self, orig_weight):
-        oft_blocks = self.oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype)
+        R = self.R
-        block_Q = oft_blocks - oft_blocks.transpose(1, 2)
+        if orig_weight.dim() == 4:
-        norm_Q = torch.norm(block_Q.flatten())
+            weight = torch.einsum("oihw, op -> pihw", orig_weight, R)
-        new_norm_Q = torch.clamp(norm_Q, max=self.constraint)
+        else:
-        block_Q = block_Q * ((new_norm_Q + 1e-8) / (norm_Q + 1e-8))
+            weight = torch.einsum("oi, op -> pi", orig_weight, R)
-        I = torch.eye(self.block_size, device=oft_blocks.device).unsqueeze(0).repeat(self.num_blocks, 1, 1)
+        updown = orig_weight @ R
-        block_R = torch.matmul(I + block_Q, (I - block_Q).inverse())
+        output_shape = [orig_weight.size(0), R.size(1)]
-
+        #output_shape = [R.size(0), orig_weight.size(1)]
-        block_R_weighted = self.oft_multiplier * block_R + (1 - self.oft_multiplier) * I
-        R = torch.block_diag(*block_R_weighted)
-        #R = self.get_weight().to(orig_weight.device, dtype=orig_weight.dtype)
-        # W = R*W_0
-        updown = orig_weight + R
-        output_shape = [R.size(0), orig_weight.size(1)]
         return self.finalize_updown(updown, orig_weight, output_shape)
     
-    # def forward(self, x, y=None):
+    def forward(self, x, y=None):
-    #     x = self.org_forward(x)
+        x = self.org_forward(x)
-    #     if self.oft_multiplier == 0.0:
+        if self.multiplier() == 0.0:
-    #         return x
+            return x
-
+        R = self.get_weight().to(x.device, dtype=x.dtype)
-    #     R = self.get_weight().to(x.device, dtype=x.dtype)
+        if x.dim() == 4:
-    #     if x.dim() == 4:
+            x = x.permute(0, 2, 3, 1)
-    #         x = x.permute(0, 2, 3, 1)
+            x = torch.matmul(x, R)
-    #         x = torch.matmul(x, R)
+            x = x.permute(0, 3, 1, 2)
-    #         x = x.permute(0, 3, 1, 2)
+        else:
-    #     else:
+            x = torch.matmul(x, R)
-    #         x = torch.matmul(x, R)
+        return x
-    #     return x

extensions-builtin/Lora/networks.py

@@ -169,6 +169,10 @@ def load_network(name, network_on_disk):
             else:
                 emb_dict[vec_name] = weight
             bundle_embeddings[emb_name] = emb_dict
+        
+        #if key_network_without_network_parts == "oft_unet":
+        #    print(key_network_without_network_parts)
+        #    pass
 
         key = convert_diffusers_name_to_compvis(key_network_without_network_parts, is_sd2)
         sd_module = shared.sd_model.network_layer_mapping.get(key, None)
@@ -185,15 +189,39 @@ def load_network(name, network_on_disk):
         elif sd_module is None and "lora_te1_text_model" in key_network_without_network_parts:
             key = key_network_without_network_parts.replace("lora_te1_text_model", "0_transformer_text_model")
             sd_module = shared.sd_model.network_layer_mapping.get(key, None)
-        elif sd_module is None and "oft_unet" in key_network_without_network_parts:
-            key = key_network_without_network_parts.replace("oft_unet", "diffusion_model")
-            sd_module = shared.sd_model.network_layer_mapping.get(key, None)
 
             # some SD1 Loras also have correct compvis keys
             if sd_module is None:
                 key = key_network_without_network_parts.replace("lora_te1_text_model", "transformer_text_model")
                 sd_module = shared.sd_model.network_layer_mapping.get(key, None)
 
+        elif sd_module is None and "oft_unet" in key_network_without_network_parts:
+        #    UNET_TARGET_REPLACE_MODULE_ALL_LINEAR = ["Transformer2DModel"]
+        #    UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 = ["ResnetBlock2D", "Downsample2D", "Upsample2D"]
+            UNET_TARGET_REPLACE_MODULE_ATTN_ONLY = ["CrossAttention"]
+            # TODO: Change matchedm odules based on whether all linear, conv, etc
+
+            key = key_network_without_network_parts.replace("oft_unet", "diffusion_model")
+            sd_module = shared.sd_model.network_layer_mapping.get(key, None)
+            #key_no_suffix = key.rsplit("_to_", 1)[0]
+            ## Match all modules of class CrossAttention
+            #replace_module_list = []
+            #for module_type in UNET_TARGET_REPLACE_MODULE_ATTN_ONLY:
+            #    replace_module_list += [module for k, module in shared.sd_model.network_layer_mapping.items() if module_type in module.__class__.__name__]
+
+            #matched_module = replace_module_list.get(key_no_suffix, None)
+            #if key.endswith('to_q'):
+            #    sd_module = matched_module.to_q or None
+            #if key.endswith('to_k'):
+            #    sd_module = matched_module.to_k or None
+            #if key.endswith('to_v'):
+            #    sd_module = matched_module.to_v or None
+            #if key.endswith('to_out_0'):
+            #    sd_module = matched_module.to_out[0] or None
+            #if key.endswith('to_out_1'):
+            #    sd_module = matched_module.to_out[1] or None
+
+
         if sd_module is None:
             keys_failed_to_match[key_network] = key
             continue
@@ -214,6 +242,14 @@ def load_network(name, network_on_disk):
             raise AssertionError(f"Could not find a module type (out of {', '.join([x.__class__.__name__ for x in module_types])}) that would accept those keys: {', '.join(weights.w)}")
 
         net.modules[key] = net_module
+    
+    # replaces forward method of original Linear
+    # applied_to_count = 0
+    #for key, created_module in net.modules.items():
+    #    if isinstance(created_module, network_oft.NetworkModuleOFT):
+    #        net_module.apply_to()
+            #applied_to_count += 1
+    # print(f'Applied OFT modules: {applied_to_count}')
 
     embeddings = {}
     for emb_name, data in bundle_embeddings.items():