diff options
Diffstat (limited to 'extensions-builtin/Lora')
| -rw-r--r-- | extensions-builtin/Lora/lyco_helpers.py | 47 | ||||
| -rw-r--r-- | extensions-builtin/Lora/network_oft.py | 82 | ||||
| -rw-r--r-- | extensions-builtin/Lora/networks.py | 16 | ||||
| -rw-r--r-- | extensions-builtin/Lora/ui_extra_networks_lora.py | 7 |
4 files changed, 149 insertions, 3 deletions
diff --git a/extensions-builtin/Lora/lyco_helpers.py b/extensions-builtin/Lora/lyco_helpers.py index 279b34bc..1679a0ce 100644 --- a/extensions-builtin/Lora/lyco_helpers.py +++ b/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
+
diff --git a/extensions-builtin/Lora/network_oft.py b/extensions-builtin/Lora/network_oft.py new file mode 100644 index 00000000..fa647020 --- /dev/null +++ b/extensions-builtin/Lora/network_oft.py @@ -0,0 +1,82 @@ +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] + + self.scale = 1.0 + + # 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(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 = 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 + output_shape = orig_weight.shape + return self.finalize_updown(updown, orig_weight, output_shape) diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index 60d8dec4..629bf853 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -11,6 +11,7 @@ import network_ia3 import network_lokr
import network_full
import network_norm
+import network_oft
import torch
from typing import Union
@@ -28,6 +29,7 @@ module_types = [ network_full.ModuleTypeFull(),
network_norm.ModuleTypeNorm(),
network_glora.ModuleTypeGLora(),
+ network_oft.ModuleTypeOFT(),
]
@@ -157,7 +159,8 @@ def load_network(name, network_on_disk): bundle_embeddings = {}
for key_network, weight in sd.items():
- key_network_without_network_parts, network_part = key_network.split(".", 1)
+ key_network_without_network_parts, _, network_part = key_network.partition(".")
+
if key_network_without_network_parts == "bundle_emb":
emb_name, vec_name = network_part.split(".", 1)
emb_dict = bundle_embeddings.get(emb_name, {})
@@ -189,6 +192,17 @@ 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
diff --git a/extensions-builtin/Lora/ui_extra_networks_lora.py b/extensions-builtin/Lora/ui_extra_networks_lora.py index 55409a78..df02c663 100644 --- a/extensions-builtin/Lora/ui_extra_networks_lora.py +++ b/extensions-builtin/Lora/ui_extra_networks_lora.py @@ -17,6 +17,8 @@ class ExtraNetworksPageLora(ui_extra_networks.ExtraNetworksPage): def create_item(self, name, index=None, enable_filter=True):
lora_on_disk = networks.available_networks.get(name)
+ if lora_on_disk is None:
+ return
path, ext = os.path.splitext(lora_on_disk.filename)
@@ -66,9 +68,10 @@ class ExtraNetworksPageLora(ui_extra_networks.ExtraNetworksPage): return item
def list_items(self):
- for index, name in enumerate(networks.available_networks):
+ # instantiate a list to protect against concurrent modification
+ names = list(networks.available_networks)
+ for index, name in enumerate(names):
item = self.create_item(name, index)
-
if item is not None:
yield item
|