Merge branch 'release_candidate'

1 files changed, 68 insertions, 32 deletions

diff --git a/extensions-builtin/Lora/network_oft.py b/extensions-builtin/Lora/network_oft.py
index fa647020..7821a8a7 100644
--- a/extensions-builtin/Lora/network_oft.py
+++ b/extensions-builtin/Lora/network_oft.py
@@ -1,6 +1,5 @@
 import torch
 import network
-from lyco_helpers import factorization
 from einops import rearrange
 
 
@@ -22,20 +21,28 @@ class NetworkModuleOFT(network.NetworkModule):
         self.org_module: list[torch.Module] = [self.sd_module]
 
         self.scale = 1.0
+        self.is_R = False
+        self.is_boft = False
 
-        # kohya-ss
+        # kohya-ss/New LyCORIS OFT/BOFT
         if "oft_blocks" in weights.w.keys():
-            self.is_kohya = True
             self.oft_blocks = weights.w["oft_blocks"] # (num_blocks, block_size, block_size)
-            self.alpha = weights.w["alpha"] # alpha is constraint
+            self.alpha = weights.w.get("alpha", None) # alpha is constraint
             self.dim = self.oft_blocks.shape[0] # lora dim
-        # LyCORIS
+        # Old LyCORIS OFT
         elif "oft_diag" in weights.w.keys():
-            self.is_kohya = False
+            self.is_R = True
             self.oft_blocks = weights.w["oft_diag"]
             # self.alpha is unused
             self.dim = self.oft_blocks.shape[1] # (num_blocks, block_size, block_size)
 
+        # LyCORIS BOFT
+        if self.oft_blocks.dim() == 4:
+            self.is_boft = True
+        self.rescale = weights.w.get('rescale', None)
+        if self.rescale is not None:
+            self.rescale = self.rescale.reshape(-1, *[1]*(self.org_module[0].weight.dim() - 1))
+
         is_linear = type(self.sd_module) in [torch.nn.Linear, torch.nn.modules.linear.NonDynamicallyQuantizableLinear]
         is_conv = type(self.sd_module) in [torch.nn.Conv2d]
         is_other_linear = type(self.sd_module) in [torch.nn.MultiheadAttention] # unsupported
@@ -47,36 +54,65 @@ class NetworkModuleOFT(network.NetworkModule):
         elif is_other_linear:
             self.out_dim = self.sd_module.embed_dim
 
-        if self.is_kohya:
-            self.constraint = self.alpha * self.out_dim
-            self.num_blocks = self.dim
-            self.block_size = self.out_dim // self.dim
-        else:
+        self.num_blocks = self.dim
+        self.block_size = self.out_dim // self.dim
+        self.constraint = (0 if self.alpha is None else self.alpha) * self.out_dim
+        if self.is_R:
             self.constraint = None
-            self.block_size, self.num_blocks = factorization(self.out_dim, self.dim)
+            self.block_size = self.dim
+            self.num_blocks = self.out_dim // self.dim
+        elif self.is_boft:
+            self.boft_m = self.oft_blocks.shape[0]
+            self.num_blocks = self.oft_blocks.shape[1]
+            self.block_size = self.oft_blocks.shape[2]
+            self.boft_b = self.block_size
 
     def calc_updown(self, orig_weight):
-        oft_blocks = self.oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype)
-        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 = self.oft_blocks.to(orig_weight.device)
+        eye = torch.eye(self.block_size, device=oft_blocks.device)
+
+        if not self.is_R:
+            block_Q = oft_blocks - oft_blocks.transpose(-1, -2) # ensure skew-symmetric orthogonal matrix
+            if self.constraint != 0:
+                norm_Q = torch.norm(block_Q.flatten())
+                new_norm_Q = torch.clamp(norm_Q, max=self.constraint.to(oft_blocks.device))
+                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)
-
-        # 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)
-        merged_weight = torch.einsum(
-            'k n m, k n ... -> k m ...',
-            R,
-            merged_weight
-        )
-        merged_weight = rearrange(merged_weight, 'k m ... -> (k m) ...')
-
-        updown = merged_weight.to(orig_weight.device, dtype=orig_weight.dtype) - orig_weight
+        R = oft_blocks.to(orig_weight.device)
+
+        if not self.is_boft:
+            # 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)
+            merged_weight = torch.einsum(
+                'k n m, k n ... -> k m ...',
+                R,
+                merged_weight
+            )
+            merged_weight = rearrange(merged_weight, 'k m ... -> (k m) ...')
+        else:
+            # TODO: determine correct value for scale
+            scale = 1.0
+            m = self.boft_m
+            b = self.boft_b
+            r_b = b // 2
+            inp = orig_weight
+            for i in range(m):
+                bi = R[i] # b_num, b_size, b_size
+                if i == 0:
+                    # Apply multiplier/scale and rescale into first weight
+                    bi = bi * scale + (1 - scale) * eye
+                inp = rearrange(inp, "(c g k) ... -> (c k g) ...", g=2, k=2**i * r_b)
+                inp = rearrange(inp, "(d b) ... -> d b ...", b=b)
+                inp = torch.einsum("b i j, b j ... -> b i ...", bi, inp)
+                inp = rearrange(inp, "d b ... -> (d b) ...")
+                inp = rearrange(inp, "(c k g) ... -> (c g k) ...", g=2, k=2**i * r_b)
+            merged_weight = inp
+
+        # Rescale mechanism
+        if self.rescale is not None:
+            merged_weight = self.rescale.to(merged_weight) * merged_weight
+
+        updown = merged_weight.to(orig_weight.device) - orig_weight.to(merged_weight.dtype)
         output_shape = orig_weight.shape
         return self.finalize_updown(updown, orig_weight, output_shape)