Support inference with OFT networks

extensions-builtin/Lora/lyco_helpers.py

@@ -19,3 +19,50 @@ def rebuild_cp_decomposition(up, down, mid):
     up = up.reshape(up.size(0), -1)
     down = down.reshape(down.size(0), -1)
     return torch.einsum('n m k l, i n, m j -> i j k l', mid, up, down)
+
+
+# copied from https://github.com/KohakuBlueleaf/LyCORIS/blob/dev/lycoris/modules/lokr.py
+def factorization(dimension: int, factor:int=-1) -> tuple[int, int]:
+    '''
+    return a tuple of two value of input dimension decomposed by the number closest to factor
+    second value is higher or equal than first value.
+
+    In LoRA with Kroneckor Product, first value is a value for weight scale.
+    secon value is a value for weight.
+
+    Becuase of non-commutative property, A⊗B ≠ B⊗A. Meaning of two matrices is slightly different.
+
+    examples)
+    factor
+        -1               2                4               8               16               ...
+    127 -> 1, 127   127 -> 1, 127    127 -> 1, 127   127 -> 1, 127   127 -> 1, 127
+    128 -> 8, 16    128 -> 2, 64     128 -> 4, 32    128 -> 8, 16    128 -> 8, 16
+    250 -> 10, 25   250 -> 2, 125    250 -> 2, 125   250 -> 5, 50    250 -> 10, 25
+    360 -> 8, 45    360 -> 2, 180    360 -> 4, 90    360 -> 8, 45    360 -> 12, 30
+    512 -> 16, 32   512 -> 2, 256    512 -> 4, 128   512 -> 8, 64    512 -> 16, 32
+    1024 -> 32, 32  1024 -> 2, 512   1024 -> 4, 256  1024 -> 8, 128  1024 -> 16, 64
+    '''
+
+    if factor > 0 and (dimension % factor) == 0:
+        m = factor
+        n = dimension // factor
+        if m > n:
+            n, m = m, n
+        return m, n
+    if factor < 0:
+        factor = dimension
+    m, n = 1, dimension
+    length = m + n
+    while m<n:
+        new_m = m + 1
+        while dimension%new_m != 0:
+            new_m += 1
+        new_n = dimension // new_m
+        if new_m + new_n > length or new_m>factor:
+            break
+        else:
+            m, n = new_m, new_n
+    if m > n:
+        n, m = m, n
+    return m, n
+

extensions-builtin/Lora/network_oft.py

@@ -0,0 +1,97 @@
+import torch
+import network
+from lyco_helpers import factorization
+from einops import rearrange
+
+
+class ModuleTypeOFT(network.ModuleType):
+    def create_module(self, net: network.Network, weights: network.NetworkWeights):
+        if all(x in weights.w for x in ["oft_blocks"]) or all(x in weights.w for x in ["oft_diag"]):
+            return NetworkModuleOFT(net, weights)
+
+        return None
+
+# Supports both kohya-ss' implementation of COFT  https://github.com/kohya-ss/sd-scripts/blob/main/networks/oft.py
+# and KohakuBlueleaf's implementation of OFT/COFT https://github.com/KohakuBlueleaf/LyCORIS/blob/dev/lycoris/modules/diag_oft.py
+class NetworkModuleOFT(network.NetworkModule):
+    def __init__(self,  net: network.Network, weights: network.NetworkWeights):
+
+        super().__init__(net, weights)
+
+        self.lin_module = None
+        self.org_module: list[torch.Module] = [self.sd_module]
+
+        # kohya-ss
+        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.dim = self.oft_blocks.shape[0] # lora dim
+        # LyCORIS
+        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)
+
+        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
+
+        if is_linear:
+            self.out_dim = self.sd_module.out_features
+        elif is_conv:
+            self.out_dim = self.sd_module.out_channels
+        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.constraint = None
+            self.block_size, self.num_blocks = factorization(self.out_dim, self.dim)
+
+    def calc_updown_kb(self, orig_weight, multiplier):
+        oft_blocks = self.oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype)
+        oft_blocks = oft_blocks - oft_blocks.transpose(1, 2) # ensure skew-symmetric orthogonal matrix
+
+        R = oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype)
+        R = R * multiplier + torch.eye(self.block_size, device=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) ...')
+
+        updown = merged_weight.to(orig_weight.device, dtype=orig_weight.dtype) - orig_weight
+        output_shape = orig_weight.shape
+        return self.finalize_updown(updown, orig_weight, output_shape)
+
+    def calc_updown(self, orig_weight):
+        # if alpha is a very small number as in coft, calc_scale() will return a almost zero number so we ignore it
+        multiplier = self.multiplier()
+        return self.calc_updown_kb(orig_weight, multiplier)
+
+    # override to remove the multiplier/scale factor; it's already multiplied in get_weight
+    def finalize_updown(self, updown, orig_weight, output_shape, ex_bias=None):
+        if self.bias is not None:
+            updown = updown.reshape(self.bias.shape)
+            updown += self.bias.to(orig_weight.device, dtype=orig_weight.dtype)
+            updown = updown.reshape(output_shape)
+
+        if len(output_shape) == 4:
+            updown = updown.reshape(output_shape)
+
+        if orig_weight.size().numel() == updown.size().numel():
+            updown = updown.reshape(orig_weight.shape)
+
+        if ex_bias is not None:
+            ex_bias = ex_bias * self.multiplier()
+
+        return updown, ex_bias

extensions-builtin/Lora/networks.py

@@ -11,6 +11,7 @@
 import network_lokr
 import network_full
 import network_norm
+import network_oft
 
 import torch
 from typing import Union
@@ -28,6 +29,7 @@
     network_full.ModuleTypeFull(),
     network_norm.ModuleTypeNorm(),
     network_glora.ModuleTypeGLora(),
+    network_oft.ModuleTypeOFT(),
 ]
 
 
@@ -189,6 +191,17 @@ def load_network(name, network_on_disk):
                 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)
 
+        # kohya_ss OFT module
+        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)
+
+        # KohakuBlueLeaf OFT module
+        if sd_module is None and "oft_diag" in key:
+            key = key_network_without_network_parts.replace("lora_unet", "diffusion_model")
+            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)
+
         if sd_module is None:
             keys_failed_to_match[key_network] = key
             continue