Merge pull request #14871 from v0xie/boft

Support inference with LyCORIS BOFT networks

extensions-builtin/Lora/network_oft.py

@@ -22,6 +22,8 @@ class NetworkModuleOFT(network.NetworkModule):
         self.org_module: list[torch.Module] = [self.sd_module]
 
         self.scale = 1.0
+        self.is_kohya = False
+        self.is_boft = False
 
         # kohya-ss
         if "oft_blocks" in weights.w.keys():
@@ -29,13 +31,19 @@ class NetworkModuleOFT(network.NetworkModule):
             self.oft_blocks = weights.w["oft_blocks"] # (num_blocks, block_size, block_size)
             self.alpha = weights.w["alpha"] # alpha is constraint
             self.dim = self.oft_blocks.shape[0] # lora dim
-        # LyCORIS
+        # LyCORIS OFT
         elif "oft_diag" in weights.w.keys():
-            self.is_kohya = False
             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 weights.w["oft_diag"].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
@@ -51,6 +59,13 @@ class NetworkModuleOFT(network.NetworkModule):
             self.constraint = self.alpha * self.out_dim
             self.num_blocks = self.dim
             self.block_size = self.out_dim // self.dim
+        elif self.is_boft:
+            self.constraint = None
+            self.boft_m = weights.w["oft_diag"].shape[0]
+            self.block_num = weights.w["oft_diag"].shape[1]
+            self.block_size = weights.w["oft_diag"].shape[2]
+            self.boft_b = self.block_size
+            #self.block_size, self.block_num = butterfly_factor(self.out_dim, self.dim)
         else:
             self.constraint = None
             self.block_size, self.num_blocks = factorization(self.out_dim, self.dim)
@@ -68,14 +83,37 @@ class NetworkModuleOFT(network.NetworkModule):
 
         R = oft_blocks.to(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)
-        merged_weight = torch.einsum(
-            'k n m, k n ... -> k m ...',
-            R,
-            merged_weight
-        )
-        merged_weight = rearrange(merged_weight, 'k m ... -> (k m) ...')
+        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