diff --git a/extensions-builtin/Lora/network_oft.py b/extensions-builtin/Lora/network_oft.py index f5f32c238..e0672ba6d 100644 --- a/extensions-builtin/Lora/network_oft.py +++ b/extensions-builtin/Lora/network_oft.py @@ -32,21 +32,27 @@ class NetworkModuleOFT(network.NetworkModule): self.org_module: list[torch.Module] = [self.sd_module] self.org_weight = self.org_module[0].weight.to(self.org_module[0].weight.device, copy=True) #self.org_weight = self.org_module[0].weight.to(devices.cpu, copy=True) - self.R = self.get_weight(self.oft_blocks) + init_multiplier = self.multiplier() * self.calc_scale() + self.last_multiplier = init_multiplier + self.R = self.get_weight(self.oft_blocks, init_multiplier) self.merged_weight = self.merge_weight() self.apply_to() self.merged = False + # weights_backup = getattr(self.org_module[0], 'network_weights_backup', None) + # if weights_backup is None: + # self.org_module[0].network_weights_backup = self.org_weight + def merge_weight(self): - org_sd = self.org_module[0].state_dict() + #org_sd = self.org_module[0].state_dict() R = self.R.to(self.org_weight.device, dtype=self.org_weight.dtype) if self.org_weight.dim() == 4: weight = torch.einsum("oihw, op -> pihw", self.org_weight, R) else: weight = torch.einsum("oi, op -> pi", self.org_weight, R) - org_sd['weight'] = weight + #org_sd['weight'] = weight # replace weight #self.org_module[0].load_state_dict(org_sd) return weight @@ -74,6 +80,7 @@ class NetworkModuleOFT(network.NetworkModule): self.org_module[0].register_forward_hook(self.forward_hook) def get_weight(self, oft_blocks, multiplier=None): + multiplier = multiplier.to(oft_blocks.device, dtype=oft_blocks.dtype) constraint = self.constraint.to(oft_blocks.device, dtype=oft_blocks.dtype) block_Q = oft_blocks - oft_blocks.transpose(1, 2) norm_Q = torch.norm(block_Q.flatten()) @@ -81,9 +88,9 @@ class NetworkModuleOFT(network.NetworkModule): block_Q = block_Q * ((new_norm_Q + 1e-8) / (norm_Q + 1e-8)) m_I = torch.eye(self.block_size, device=oft_blocks.device).unsqueeze(0).repeat(self.num_blocks, 1, 1) block_R = torch.matmul(m_I + block_Q, (m_I - block_Q).inverse()) - #block_R_weighted = multiplier * block_R + (1 - multiplier) * I - #R = torch.block_diag(*block_R_weighted) - R = torch.block_diag(*block_R) + block_R_weighted = multiplier * block_R + (1 - multiplier) * m_I + R = torch.block_diag(*block_R_weighted) + #R = torch.block_diag(*block_R) return R @@ -93,6 +100,8 @@ class NetworkModuleOFT(network.NetworkModule): #R = self.R.to(orig_weight.device, dtype=orig_weight.dtype) ##self.R = R + #R = self.R.to(orig_weight.device, dtype=orig_weight.dtype) + ##self.R = R #if orig_weight.dim() == 4: # weight = torch.einsum("oihw, op -> pihw", orig_weight, R) #else: @@ -103,19 +112,33 @@ class NetworkModuleOFT(network.NetworkModule): updown = torch.zeros_like(orig_weight, device=orig_weight.device, dtype=orig_weight.dtype) #updown = orig_weight output_shape = orig_weight.shape - #orig_weight = self.merged_weight.to(orig_weight.device, dtype=orig_weight.dtype) + orig_weight = self.merged_weight.to(orig_weight.device, dtype=orig_weight.dtype) #output_shape = self.oft_blocks.shape return self.finalize_updown(updown, orig_weight, output_shape) def pre_forward_hook(self, module, input): - if not self.merged: + multiplier = self.multiplier() * self.calc_scale() + if not multiplier==self.last_multiplier or not self.merged: + + #if multiplier != self.last_multiplier or not self.merged: + self.R = self.get_weight(self.oft_blocks, multiplier) + self.last_multiplier = multiplier + self.merged_weight = self.merge_weight() self.replace_weight(self.merged_weight) + #elif not self.merged: + # self.replace_weight(self.merged_weight) def forward_hook(self, module, args, output): - if self.merged: - pass + pass + #output = output * self.multiplier() * self.calc_scale() + #if len(args) > 0: + # y = args[0] + # output = output + y + #return output + #if self.merged: + # pass #self.restore_weight() #print(f'Forward hook in {self.network_key} called')