fix: return orig weights during updown, merge weights before forward

1 files changed, 69 insertions, 21 deletions

diff --git a/extensions-builtin/Lora/network_oft.py b/extensions-builtin/Lora/network_oft.py
index 8e561ab0..f5f32c23 100644
--- a/extensions-builtin/Lora/network_oft.py
+++ b/extensions-builtin/Lora/network_oft.py
@@ -1,5 +1,6 @@
 import torch
 import network
+from modules import devices
 
 
 class ModuleTypeOFT(network.ModuleType):
@@ -29,23 +30,56 @@ class NetworkModuleOFT(network.NetworkModule):
         self.block_size = self.out_dim // self.num_blocks
 
         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)
+
+        self.merged_weight = self.merge_weight()
         self.apply_to()
+        self.merged = False
+
+
+    def merge_weight(self):
+        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
+        # replace weight
+        #self.org_module[0].load_state_dict(org_sd)
+        return weight
+        pass
+    
+    def replace_weight(self, new_weight):
+        org_sd = self.org_module[0].state_dict()
+        org_sd['weight'] = new_weight
+        self.org_module[0].load_state_dict(org_sd)
+        self.merged = True
+
+    def restore_weight(self):
+        org_sd = self.org_module[0].state_dict()
+        org_sd['weight'] = self.org_weight
+        self.org_module[0].load_state_dict(org_sd)
+        self.merged = False
+
 
     # replace forward method of original linear rather than replacing the module
     # how do we revert this to unload the weights?
     def apply_to(self):
         self.org_forward = self.org_module[0].forward
         #self.org_module[0].forward = self.forward
+        self.org_module[0].register_forward_pre_hook(self.pre_forward_hook)
         self.org_module[0].register_forward_hook(self.forward_hook)
 
     def get_weight(self, oft_blocks, multiplier=None):
-        self.constraint = self.constraint.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())
-        new_norm_Q = torch.clamp(norm_Q, max=self.constraint)
+        new_norm_Q = torch.clamp(norm_Q, max=constraint)
         block_Q = block_Q * ((new_norm_Q + 1e-8) / (norm_Q + 1e-8))
-        m_I = torch.eye(self.block_size, device=self.oft_blocks.device).unsqueeze(0).repeat(self.num_blocks, 1, 1)
+        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)
@@ -54,33 +88,47 @@ class NetworkModuleOFT(network.NetworkModule):
         return R
 
     def calc_updown(self, orig_weight):
-        oft_blocks = self.oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype)
+        #oft_blocks = self.oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype)
 
-        R = self.get_weight(oft_blocks)
-        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:
-        #     weight = torch.einsum("oi, op -> pi", orig_weight, R)
+        #if orig_weight.dim() == 4:
+        #    weight = torch.einsum("oihw, op -> pihw", orig_weight, R)
+        #else:
+        #    weight = torch.einsum("oi, op -> pi", orig_weight, R)
 
-        updown = orig_weight @ R
-        output_shape = self.oft_blocks.shape
+        #updown = orig_weight @ R
+        #updown = weight
+        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)
+        #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:
+            self.replace_weight(self.merged_weight)
+
+    
     def forward_hook(self, module, args, output):
+        if self.merged:
+            pass
+            #self.restore_weight()
         #print(f'Forward hook in {self.network_key} called')
-        x = output
-        R = self.R.to(x.device, dtype=x.dtype)
 
-        if x.dim() == 4:
-            x = x.permute(0, 2, 3, 1)
-            x = torch.matmul(x, R)
-            x = x.permute(0, 3, 1, 2)
-        else:
-            x = torch.matmul(x, R)
-        return x
+        #x = output
+        #R = self.R.to(x.device, dtype=x.dtype)
+
+        #if x.dim() == 4:
+        #    x = x.permute(0, 2, 3, 1)
+        #    x = torch.matmul(x, R)
+        #    x = x.permute(0, 3, 1, 2)
+        #else:
+        #    x = torch.matmul(x, R)
+        #return x
 
     # def forward(self, x, y=None):
     #     x = self.org_forward(x)