From ec718f76b58b183859ed732e11ec748c41a13f76 Mon Sep 17 00:00:00 2001 From: v0xie <28695009+v0xie@users.noreply.github.com> Date: Tue, 17 Oct 2023 23:35:50 -0700 Subject: wip incorrect OFT implementation --- extensions-builtin/Lora/networks.py | 5 +++++ 1 file changed, 5 insertions(+) (limited to 'extensions-builtin/Lora/networks.py') diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index 60d8dec4..bd1f1b75 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(), ] @@ -183,6 +185,9 @@ def load_network(name, network_on_disk): elif sd_module is None and "lora_te1_text_model" in key_network_without_network_parts: 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) + 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) # some SD1 Loras also have correct compvis keys if sd_module is None: -- cgit v1.2.3 From 1c6efdbba774d603c592debaccd6f5ad827bd1b2 Mon Sep 17 00:00:00 2001 From: v0xie <28695009+v0xie@users.noreply.github.com> Date: Wed, 18 Oct 2023 04:16:01 -0700 Subject: inference working but SLOW --- extensions-builtin/Lora/network_oft.py | 73 +++++++++++++++++----------------- extensions-builtin/Lora/networks.py | 42 +++++++++++++++++-- 2 files changed, 75 insertions(+), 40 deletions(-) (limited to 'extensions-builtin/Lora/networks.py') diff --git a/extensions-builtin/Lora/network_oft.py b/extensions-builtin/Lora/network_oft.py index 9ddb175c..f085eca5 100644 --- a/extensions-builtin/Lora/network_oft.py +++ b/extensions-builtin/Lora/network_oft.py @@ -12,6 +12,7 @@ class ModuleTypeOFT(network.ModuleType): # adapted from https://github.com/kohya-ss/sd-scripts/blob/main/networks/oft.py class NetworkModuleOFT(network.NetworkModule): def __init__(self, net: network.Network, weights: network.NetworkWeights): + super().__init__(net, weights) self.oft_blocks = weights.w["oft_blocks"] @@ -20,24 +21,29 @@ class NetworkModuleOFT(network.NetworkModule): self.dim = self.oft_blocks.shape[0] self.num_blocks = self.dim - #if type(self.alpha) == torch.Tensor: - # self.alpha = self.alpha.detach().numpy() - if "Linear" in self.sd_module.__class__.__name__: self.out_dim = self.sd_module.out_features elif "Conv" in self.sd_module.__class__.__name__: self.out_dim = self.sd_module.out_channels - self.constraint = self.alpha * self.out_dim + self.constraint = self.alpha + #self.constraint = self.alpha * self.out_dim self.block_size = self.out_dim // self.num_blocks - self.oft_multiplier = self.multiplier() + self.org_module: list[torch.Module] = [self.sd_module] + + self.R = self.get_weight() - # replace forward method of original linear rather than replacing the module - # self.org_forward = self.sd_module.forward - # self.sd_module.forward = self.forward + self.apply_to() + + # replace forward method of original linear rather than replacing the module + def apply_to(self): + self.org_forward = self.org_module[0].forward + self.org_module[0].forward = self.forward - def get_weight(self): + def get_weight(self, multiplier=None): + if not multiplier: + multiplier = self.multiplier() block_Q = self.oft_blocks - self.oft_blocks.transpose(1, 2) norm_Q = torch.norm(block_Q.flatten()) new_norm_Q = torch.clamp(norm_Q, max=self.constraint) @@ -45,38 +51,31 @@ class NetworkModuleOFT(network.NetworkModule): I = torch.eye(self.block_size, device=self.oft_blocks.device).unsqueeze(0).repeat(self.num_blocks, 1, 1) block_R = torch.matmul(I + block_Q, (I - block_Q).inverse()) - block_R_weighted = self.oft_multiplier * block_R + (1 - self.oft_multiplier) * I + block_R_weighted = multiplier * block_R + (1 - multiplier) * I R = torch.block_diag(*block_R_weighted) return R def calc_updown(self, orig_weight): - oft_blocks = self.oft_blocks.to(orig_weight.device, dtype=orig_weight.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) - block_Q = block_Q * ((new_norm_Q + 1e-8) / (norm_Q + 1e-8)) - I = torch.eye(self.block_size, device=oft_blocks.device).unsqueeze(0).repeat(self.num_blocks, 1, 1) - block_R = torch.matmul(I + block_Q, (I - block_Q).inverse()) - - block_R_weighted = self.oft_multiplier * block_R + (1 - self.oft_multiplier) * I - R = torch.block_diag(*block_R_weighted) - #R = self.get_weight().to(orig_weight.device, dtype=orig_weight.dtype) - # W = R*W_0 - updown = orig_weight + R - output_shape = [R.size(0), orig_weight.size(1)] + R = self.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 = [orig_weight.size(0), R.size(1)] + #output_shape = [R.size(0), orig_weight.size(1)] return self.finalize_updown(updown, orig_weight, output_shape) - # def forward(self, x, y=None): - # x = self.org_forward(x) - # if self.oft_multiplier == 0.0: - # return x - - # R = self.get_weight().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) + if self.multiplier() == 0.0: + return x + R = self.get_weight().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 diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index bd1f1b75..e5e73450 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -169,6 +169,10 @@ def load_network(name, network_on_disk): else: emb_dict[vec_name] = weight bundle_embeddings[emb_name] = emb_dict + + #if key_network_without_network_parts == "oft_unet": + # print(key_network_without_network_parts) + # pass key = convert_diffusers_name_to_compvis(key_network_without_network_parts, is_sd2) sd_module = shared.sd_model.network_layer_mapping.get(key, None) @@ -185,15 +189,39 @@ def load_network(name, network_on_disk): elif sd_module is None and "lora_te1_text_model" in key_network_without_network_parts: 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) - 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) # some SD1 Loras also have correct compvis keys if sd_module is None: 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) + elif sd_module is None and "oft_unet" in key_network_without_network_parts: + # UNET_TARGET_REPLACE_MODULE_ALL_LINEAR = ["Transformer2DModel"] + # UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 = ["ResnetBlock2D", "Downsample2D", "Upsample2D"] + UNET_TARGET_REPLACE_MODULE_ATTN_ONLY = ["CrossAttention"] + # TODO: Change matchedm odules based on whether all linear, conv, etc + + key = key_network_without_network_parts.replace("oft_unet", "diffusion_model") + sd_module = shared.sd_model.network_layer_mapping.get(key, None) + #key_no_suffix = key.rsplit("_to_", 1)[0] + ## Match all modules of class CrossAttention + #replace_module_list = [] + #for module_type in UNET_TARGET_REPLACE_MODULE_ATTN_ONLY: + # replace_module_list += [module for k, module in shared.sd_model.network_layer_mapping.items() if module_type in module.__class__.__name__] + + #matched_module = replace_module_list.get(key_no_suffix, None) + #if key.endswith('to_q'): + # sd_module = matched_module.to_q or None + #if key.endswith('to_k'): + # sd_module = matched_module.to_k or None + #if key.endswith('to_v'): + # sd_module = matched_module.to_v or None + #if key.endswith('to_out_0'): + # sd_module = matched_module.to_out[0] or None + #if key.endswith('to_out_1'): + # sd_module = matched_module.to_out[1] or None + + if sd_module is None: keys_failed_to_match[key_network] = key continue @@ -214,6 +242,14 @@ def load_network(name, network_on_disk): raise AssertionError(f"Could not find a module type (out of {', '.join([x.__class__.__name__ for x in module_types])}) that would accept those keys: {', '.join(weights.w)}") net.modules[key] = net_module + + # replaces forward method of original Linear + # applied_to_count = 0 + #for key, created_module in net.modules.items(): + # if isinstance(created_module, network_oft.NetworkModuleOFT): + # net_module.apply_to() + #applied_to_count += 1 + # print(f'Applied OFT modules: {applied_to_count}') embeddings = {} for emb_name, data in bundle_embeddings.items(): -- cgit v1.2.3 From d10c4db57ed08234a7aed5f530f269ff78544ab0 Mon Sep 17 00:00:00 2001 From: v0xie <28695009+v0xie@users.noreply.github.com> Date: Thu, 19 Oct 2023 12:52:14 -0700 Subject: style: formatting --- extensions-builtin/Lora/network_oft.py | 4 ++-- extensions-builtin/Lora/networks.py | 35 ---------------------------------- 2 files changed, 2 insertions(+), 37 deletions(-) (limited to 'extensions-builtin/Lora/networks.py') diff --git a/extensions-builtin/Lora/network_oft.py b/extensions-builtin/Lora/network_oft.py index 2af1bc4c..0a87958e 100644 --- a/extensions-builtin/Lora/network_oft.py +++ b/extensions-builtin/Lora/network_oft.py @@ -37,7 +37,7 @@ class NetworkModuleOFT(network.NetworkModule): def apply_to(self): self.org_forward = self.org_module[0].forward self.org_module[0].forward = self.forward - + def get_weight(self, oft_blocks, multiplier=None): block_Q = oft_blocks - oft_blocks.transpose(1, 2) norm_Q = torch.norm(block_Q.flatten()) @@ -66,7 +66,7 @@ class NetworkModuleOFT(network.NetworkModule): output_shape = self.oft_blocks.shape return self.finalize_updown(updown, orig_weight, output_shape) - + def forward(self, x, y=None): x = self.org_forward(x) if self.multiplier() == 0.0: diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index e5e73450..78a97033 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -169,10 +169,6 @@ def load_network(name, network_on_disk): else: emb_dict[vec_name] = weight bundle_embeddings[emb_name] = emb_dict - - #if key_network_without_network_parts == "oft_unet": - # print(key_network_without_network_parts) - # pass key = convert_diffusers_name_to_compvis(key_network_without_network_parts, is_sd2) sd_module = shared.sd_model.network_layer_mapping.get(key, None) @@ -196,31 +192,8 @@ def load_network(name, network_on_disk): sd_module = shared.sd_model.network_layer_mapping.get(key, None) elif sd_module is None and "oft_unet" in key_network_without_network_parts: - # UNET_TARGET_REPLACE_MODULE_ALL_LINEAR = ["Transformer2DModel"] - # UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 = ["ResnetBlock2D", "Downsample2D", "Upsample2D"] - UNET_TARGET_REPLACE_MODULE_ATTN_ONLY = ["CrossAttention"] - # TODO: Change matchedm odules based on whether all linear, conv, etc - key = key_network_without_network_parts.replace("oft_unet", "diffusion_model") sd_module = shared.sd_model.network_layer_mapping.get(key, None) - #key_no_suffix = key.rsplit("_to_", 1)[0] - ## Match all modules of class CrossAttention - #replace_module_list = [] - #for module_type in UNET_TARGET_REPLACE_MODULE_ATTN_ONLY: - # replace_module_list += [module for k, module in shared.sd_model.network_layer_mapping.items() if module_type in module.__class__.__name__] - - #matched_module = replace_module_list.get(key_no_suffix, None) - #if key.endswith('to_q'): - # sd_module = matched_module.to_q or None - #if key.endswith('to_k'): - # sd_module = matched_module.to_k or None - #if key.endswith('to_v'): - # sd_module = matched_module.to_v or None - #if key.endswith('to_out_0'): - # sd_module = matched_module.to_out[0] or None - #if key.endswith('to_out_1'): - # sd_module = matched_module.to_out[1] or None - if sd_module is None: keys_failed_to_match[key_network] = key @@ -242,14 +215,6 @@ def load_network(name, network_on_disk): raise AssertionError(f"Could not find a module type (out of {', '.join([x.__class__.__name__ for x in module_types])}) that would accept those keys: {', '.join(weights.w)}") net.modules[key] = net_module - - # replaces forward method of original Linear - # applied_to_count = 0 - #for key, created_module in net.modules.items(): - # if isinstance(created_module, network_oft.NetworkModuleOFT): - # net_module.apply_to() - #applied_to_count += 1 - # print(f'Applied OFT modules: {applied_to_count}') embeddings = {} for emb_name, data in bundle_embeddings.items(): -- cgit v1.2.3 From 65ccd6305fcf72347d5ed68f03095dced865ef6e Mon Sep 17 00:00:00 2001 From: v0xie <28695009+v0xie@users.noreply.github.com> Date: Thu, 2 Nov 2023 00:11:32 -0700 Subject: detect diag_oft type --- extensions-builtin/Lora/networks.py | 7 +++++++ 1 file changed, 7 insertions(+) (limited to 'extensions-builtin/Lora/networks.py') diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index 78a97033..7f814706 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -191,10 +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 -- cgit v1.2.3 From 2be85f8fe03533bf3b1ad562ea58ee8227ba3b99 Mon Sep 17 00:00:00 2001 From: AUTOMATIC1111 <16777216c@gmail.com> Date: Thu, 14 Dec 2023 10:08:03 +0300 Subject: Merge pull request #14237 from ReneKroon/dev #13354 : solve lora loading issue --- extensions-builtin/Lora/networks.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) (limited to 'extensions-builtin/Lora/networks.py') diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index 7f814706..629bf853 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -159,7 +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, {}) -- cgit v1.2.3