lora extension rework to include other types of networks

extensions-builtin/Lora/extra_networks_lora.py

 from modules import extra_networks, shared
-import lora
+import networks
 
 
 class ExtraNetworkLora(extra_networks.ExtraNetwork):
@@ -9,7 +9,7 @@ class ExtraNetworkLora(extra_networks.ExtraNetwork):
     def activate(self, p, params_list):
         additional = shared.opts.sd_lora
 
-        if additional != "None" and additional in lora.available_loras and not any(x for x in params_list if x.items[0] == additional):
+        if additional != "None" and additional in networks.available_networks and not any(x for x in params_list if x.items[0] == additional):
             p.all_prompts = [x + f"<lora:{additional}:{shared.opts.extra_networks_default_multiplier}>" for x in p.all_prompts]
             params_list.append(extra_networks.ExtraNetworkParams(items=[additional, shared.opts.extra_networks_default_multiplier]))
 
@@ -21,12 +21,12 @@ class ExtraNetworkLora(extra_networks.ExtraNetwork):
             names.append(params.items[0])
             multipliers.append(float(params.items[1]) if len(params.items) > 1 else 1.0)
 
-        lora.load_loras(names, multipliers)
+        networks.load_networks(names, multipliers)
 
         if shared.opts.lora_add_hashes_to_infotext:
-            lora_hashes = []
-            for item in lora.loaded_loras:
-                shorthash = item.lora_on_disk.shorthash
+            network_hashes = []
+            for item in networks.loaded_networks:
+                shorthash = item.network_on_disk.shorthash
                 if not shorthash:
                     continue
 
@@ -36,10 +36,10 @@ class ExtraNetworkLora(extra_networks.ExtraNetwork):
 
                 alias = alias.replace(":", "").replace(",", "")
 
-                lora_hashes.append(f"{alias}: {shorthash}")
+                network_hashes.append(f"{alias}: {shorthash}")
 
-            if lora_hashes:
-                p.extra_generation_params["Lora hashes"] = ", ".join(lora_hashes)
+            if network_hashes:
+                p.extra_generation_params["Lora hashes"] = ", ".join(network_hashes)
 
     def deactivate(self, p):
         pass

extensions-builtin/Lora/lyco_helpers.py

+import torch
+
+
+def make_weight_cp(t, wa, wb):
+    temp = torch.einsum('i j k l, j r -> i r k l', t, wb)
+    return torch.einsum('i j k l, i r -> r j k l', temp, wa)
+
+
+def rebuild_conventional(up, down, shape, dyn_dim=None):
+    up = up.reshape(up.size(0), -1)
+    down = down.reshape(down.size(0), -1)
+    if dyn_dim is not None:
+        up = up[:, :dyn_dim]
+        down = down[:dyn_dim, :]
+    return (up @ down).reshape(shape)

extensions-builtin/Lora/network.py

+import os
+from collections import namedtuple
+
+import torch
+
+from modules import devices, sd_models, cache, errors, hashes, shared
+
+NetworkWeights = namedtuple('NetworkWeights', ['network_key', 'sd_key', 'w', 'sd_module'])
+
+metadata_tags_order = {"ss_sd_model_name": 1, "ss_resolution": 2, "ss_clip_skip": 3, "ss_num_train_images": 10, "ss_tag_frequency": 20}
+
+
+class NetworkOnDisk:
+    def __init__(self, name, filename):
+        self.name = name
+        self.filename = filename
+        self.metadata = {}
+        self.is_safetensors = os.path.splitext(filename)[1].lower() == ".safetensors"
+
+        def read_metadata():
+            metadata = sd_models.read_metadata_from_safetensors(filename)
+            metadata.pop('ssmd_cover_images', None)  # those are cover images, and they are too big to display in UI as text
+
+            return metadata
+
+        if self.is_safetensors:
+            try:
+                self.metadata = cache.cached_data_for_file('safetensors-metadata', "lora/" + self.name, filename, read_metadata)
+            except Exception as e:
+                errors.display(e, f"reading lora {filename}")
+
+        if self.metadata:
+            m = {}
+            for k, v in sorted(self.metadata.items(), key=lambda x: metadata_tags_order.get(x[0], 999)):
+                m[k] = v
+
+            self.metadata = m
+
+        self.alias = self.metadata.get('ss_output_name', self.name)
+
+        self.hash = None
+        self.shorthash = None
+        self.set_hash(
+            self.metadata.get('sshs_model_hash') or
+            hashes.sha256_from_cache(self.filename, "lora/" + self.name, use_addnet_hash=self.is_safetensors) or
+            ''
+        )
+
+    def set_hash(self, v):
+        self.hash = v
+        self.shorthash = self.hash[0:12]
+
+        if self.shorthash:
+            import networks
+            networks.available_network_hash_lookup[self.shorthash] = self
+
+    def read_hash(self):
+        if not self.hash:
+            self.set_hash(hashes.sha256(self.filename, "lora/" + self.name, use_addnet_hash=self.is_safetensors) or '')
+
+    def get_alias(self):
+        import networks
+        if shared.opts.lora_preferred_name == "Filename" or self.alias.lower() in networks.forbidden_network_aliases:
+            return self.name
+        else:
+            return self.alias
+
+
+class Network:  # LoraModule
+    def __init__(self, name, network_on_disk: NetworkOnDisk):
+        self.name = name
+        self.network_on_disk = network_on_disk
+        self.multiplier = 1.0
+        self.modules = {}
+        self.mtime = None
+
+        self.mentioned_name = None
+        """the text that was used to add the network to prompt - can be either name or an alias"""
+
+
+class ModuleType:
+    def create_module(self, net: Network, weights: NetworkWeights) -> Network | None:
+        return None
+
+
+class NetworkModule:
+    def __init__(self, net: Network, weights: NetworkWeights):
+        self.network = net
+        self.network_key = weights.network_key
+        self.sd_key = weights.sd_key
+        self.sd_module = weights.sd_module
+
+    def calc_updown(self, target):
+        raise NotImplementedError()
+
+    def forward(self, x, y):
+        raise NotImplementedError()
+

extensions-builtin/Lora/network_hada.py

+import lyco_helpers
+import network
+import network_lyco
+
+
+class ModuleTypeHada(network.ModuleType):
+    def create_module(self, net: network.Network, weights: network.NetworkWeights):
+        if all(x in weights.w for x in ["hada_w1_a", "hada_w1_b", "hada_w2_a", "hada_w2_b"]):
+            return NetworkModuleHada(net, weights)
+
+        return None
+
+
+class NetworkModuleHada(network_lyco.NetworkModuleLyco):
+    def __init__(self,  net: network.Network, weights: network.NetworkWeights):
+        super().__init__(net, weights)
+
+        if hasattr(self.sd_module, 'weight'):
+            self.shape = self.sd_module.weight.shape
+
+        self.w1a = weights.w["hada_w1_a"]
+        self.w1b = weights.w["hada_w1_b"]
+        self.dim = self.w1b.shape[0]
+        self.w2a = weights.w["hada_w2_a"]
+        self.w2b = weights.w["hada_w2_b"]
+
+        self.t1 = weights.w.get("hada_t1")
+        self.t2 = weights.w.get("hada_t2")
+
+        self.alpha = weights.w["alpha"].item() if "alpha" in weights.w else None
+        self.scale = weights.w["scale"].item() if "scale" in weights.w else None
+
+    def calc_updown(self, orig_weight):
+        w1a = self.w1a.to(orig_weight.device, dtype=orig_weight.dtype)
+        w1b = self.w1b.to(orig_weight.device, dtype=orig_weight.dtype)
+        w2a = self.w2a.to(orig_weight.device, dtype=orig_weight.dtype)
+        w2b = self.w2b.to(orig_weight.device, dtype=orig_weight.dtype)
+
+        output_shape = [w1a.size(0), w1b.size(1)]
+
+        if self.t1 is not None:
+            output_shape = [w1a.size(1), w1b.size(1)]
+            t1 = self.t1.to(orig_weight.device, dtype=orig_weight.dtype)
+            updown1 = lyco_helpers.make_weight_cp(t1, w1a, w1b)
+            output_shape += t1.shape[2:]
+        else:
+            if len(w1b.shape) == 4:
+                output_shape += w1b.shape[2:]
+            updown1 = lyco_helpers.rebuild_conventional(w1a, w1b, output_shape)
+
+        if self.t2 is not None:
+            t2 = self.t2.to(orig_weight.device, dtype=orig_weight.dtype)
+            updown2 = lyco_helpers.make_weight_cp(t2, w2a, w2b)
+        else:
+            updown2 = lyco_helpers.rebuild_conventional(w2a, w2b, output_shape)
+
+        updown = updown1 * updown2
+
+        return self.finalize_updown(updown, orig_weight, output_shape)

extensions-builtin/Lora/network_lora.py

+import torch
+
+import network
+from modules import devices
+
+
+class ModuleTypeLora(network.ModuleType):
+    def create_module(self, net: network.Network, weights: network.NetworkWeights):
+        if all(x in weights.w for x in ["lora_up.weight", "lora_down.weight"]):
+            return NetworkModuleLora(net, weights)
+
+        return None
+
+
+class NetworkModuleLora(network.NetworkModule):
+    def __init__(self,  net: network.Network, weights: network.NetworkWeights):
+        super().__init__(net, weights)
+
+        self.up = self.create_module(weights.w["lora_up.weight"])
+        self.down = self.create_module(weights.w["lora_down.weight"])
+        self.alpha = weights.w["alpha"] if "alpha" in weights.w else None
+
+    def create_module(self, weight, none_ok=False):
+        if weight is None and none_ok:
+            return None
+
+        if type(self.sd_module) == torch.nn.Linear:
+            module = torch.nn.Linear(weight.shape[1], weight.shape[0], bias=False)
+        elif type(self.sd_module) == torch.nn.modules.linear.NonDynamicallyQuantizableLinear:
+            module = torch.nn.Linear(weight.shape[1], weight.shape[0], bias=False)
+        elif type(self.sd_module) == torch.nn.MultiheadAttention:
+            module = torch.nn.Linear(weight.shape[1], weight.shape[0], bias=False)
+        elif type(self.sd_module) == torch.nn.Conv2d and weight.shape[2:] == (1, 1):
+            module = torch.nn.Conv2d(weight.shape[1], weight.shape[0], (1, 1), bias=False)
+        elif type(self.sd_module) == torch.nn.Conv2d and weight.shape[2:] == (3, 3):
+            module = torch.nn.Conv2d(weight.shape[1], weight.shape[0], (3, 3), bias=False)
+        else:
+            print(f'Network layer {self.network_key} matched a layer with unsupported type: {type(self.sd_module).__name__}')
+            return None
+
+        with torch.no_grad():
+            module.weight.copy_(weight)
+
+        module.to(device=devices.cpu, dtype=devices.dtype)
+        module.weight.requires_grad_(False)
+
+        return module
+
+    def calc_updown(self, target):
+        up = self.up.weight.to(target.device, dtype=target.dtype)
+        down = self.down.weight.to(target.device, dtype=target.dtype)
+
+        if up.shape[2:] == (1, 1) and down.shape[2:] == (1, 1):
+            updown = (up.squeeze(2).squeeze(2) @ down.squeeze(2).squeeze(2)).unsqueeze(2).unsqueeze(3)
+        elif up.shape[2:] == (3, 3) or down.shape[2:] == (3, 3):
+            updown = torch.nn.functional.conv2d(down.permute(1, 0, 2, 3), up).permute(1, 0, 2, 3)
+        else:
+            updown = up @ down
+
+        updown = updown * self.network.multiplier * (self.alpha / self.up.weight.shape[1] if self.alpha else 1.0)
+
+        return updown
+
+    def forward(self, x, y):
+        self.up.to(device=devices.device)
+        self.down.to(device=devices.device)
+
+        return y + self.up(self.down(x)) * self.network.multiplier * (self.alpha / self.up.weight.shape[1] if self.alpha else 1.0)
+
+

extensions-builtin/Lora/network_lyco.py

+import torch
+
+import lyco_helpers
+import network
+from modules import devices
+
+
+class NetworkModuleLyco(network.NetworkModule):
+    def __init__(self, net: network.Network, weights: network.NetworkWeights):
+        super().__init__(net, weights)
+
+        if hasattr(self.sd_module, 'weight'):
+            self.shape = self.sd_module.weight.shape
+
+        self.dim = None
+        self.bias = weights.w.get("bias")
+        self.alpha = weights.w["alpha"].item() if "alpha" in weights.w else None
+        self.scale = weights.w["scale"].item() if "scale" in weights.w else None
+
+    def finalize_updown(self, updown, orig_weight, output_shape):
+        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)
+
+        scale = (
+            self.scale if self.scale is not None
+            else self.alpha / self.dim if self.dim is not None and self.alpha is not None
+            else 1.0
+        )
+
+        return updown * scale * self.network.multiplier
+

extensions-builtin/Lora/scripts/lora_script.py

@@ -4,18 +4,19 @@ import torch
 import gradio as gr
 from fastapi import FastAPI
 
-import lora
+import network
+import networks
 import extra_networks_lora
 import ui_extra_networks_lora
 from modules import script_callbacks, ui_extra_networks, extra_networks, shared
 
 def unload():
-    torch.nn.Linear.forward = torch.nn.Linear_forward_before_lora
-    torch.nn.Linear._load_from_state_dict = torch.nn.Linear_load_state_dict_before_lora
-    torch.nn.Conv2d.forward = torch.nn.Conv2d_forward_before_lora
-    torch.nn.Conv2d._load_from_state_dict = torch.nn.Conv2d_load_state_dict_before_lora
-    torch.nn.MultiheadAttention.forward = torch.nn.MultiheadAttention_forward_before_lora
-    torch.nn.MultiheadAttention._load_from_state_dict = torch.nn.MultiheadAttention_load_state_dict_before_lora
+    torch.nn.Linear.forward = torch.nn.Linear_forward_before_network
+    torch.nn.Linear._load_from_state_dict = torch.nn.Linear_load_state_dict_before_network
+    torch.nn.Conv2d.forward = torch.nn.Conv2d_forward_before_network
+    torch.nn.Conv2d._load_from_state_dict = torch.nn.Conv2d_load_state_dict_before_network
+    torch.nn.MultiheadAttention.forward = torch.nn.MultiheadAttention_forward_before_network
+    torch.nn.MultiheadAttention._load_from_state_dict = torch.nn.MultiheadAttention_load_state_dict_before_network
 
 
 def before_ui():
@@ -23,50 +24,50 @@ def before_ui():
     extra_networks.register_extra_network(extra_networks_lora.ExtraNetworkLora())
 
 
-if not hasattr(torch.nn, 'Linear_forward_before_lora'):
-    torch.nn.Linear_forward_before_lora = torch.nn.Linear.forward
+if not hasattr(torch.nn, 'Linear_forward_before_network'):
+    torch.nn.Linear_forward_before_network = torch.nn.Linear.forward
 
-if not hasattr(torch.nn, 'Linear_load_state_dict_before_lora'):
-    torch.nn.Linear_load_state_dict_before_lora = torch.nn.Linear._load_from_state_dict
+if not hasattr(torch.nn, 'Linear_load_state_dict_before_network'):
+    torch.nn.Linear_load_state_dict_before_network = torch.nn.Linear._load_from_state_dict
 
-if not hasattr(torch.nn, 'Conv2d_forward_before_lora'):
-    torch.nn.Conv2d_forward_before_lora = torch.nn.Conv2d.forward
+if not hasattr(torch.nn, 'Conv2d_forward_before_network'):
+    torch.nn.Conv2d_forward_before_network = torch.nn.Conv2d.forward
 
-if not hasattr(torch.nn, 'Conv2d_load_state_dict_before_lora'):
-    torch.nn.Conv2d_load_state_dict_before_lora = torch.nn.Conv2d._load_from_state_dict
+if not hasattr(torch.nn, 'Conv2d_load_state_dict_before_network'):
+    torch.nn.Conv2d_load_state_dict_before_network = torch.nn.Conv2d._load_from_state_dict
 
-if not hasattr(torch.nn, 'MultiheadAttention_forward_before_lora'):
-    torch.nn.MultiheadAttention_forward_before_lora = torch.nn.MultiheadAttention.forward
+if not hasattr(torch.nn, 'MultiheadAttention_forward_before_network'):
+    torch.nn.MultiheadAttention_forward_before_network = torch.nn.MultiheadAttention.forward
 
-if not hasattr(torch.nn, 'MultiheadAttention_load_state_dict_before_lora'):
-    torch.nn.MultiheadAttention_load_state_dict_before_lora = torch.nn.MultiheadAttention._load_from_state_dict
+if not hasattr(torch.nn, 'MultiheadAttention_load_state_dict_before_network'):
+    torch.nn.MultiheadAttention_load_state_dict_before_network = torch.nn.MultiheadAttention._load_from_state_dict
 
-torch.nn.Linear.forward = lora.lora_Linear_forward
-torch.nn.Linear._load_from_state_dict = lora.lora_Linear_load_state_dict
-torch.nn.Conv2d.forward = lora.lora_Conv2d_forward
-torch.nn.Conv2d._load_from_state_dict = lora.lora_Conv2d_load_state_dict
-torch.nn.MultiheadAttention.forward = lora.lora_MultiheadAttention_forward
-torch.nn.MultiheadAttention._load_from_state_dict = lora.lora_MultiheadAttention_load_state_dict
+torch.nn.Linear.forward = networks.network_Linear_forward
+torch.nn.Linear._load_from_state_dict = networks.network_Linear_load_state_dict
+torch.nn.Conv2d.forward = networks.network_Conv2d_forward
+torch.nn.Conv2d._load_from_state_dict = networks.network_Conv2d_load_state_dict
+torch.nn.MultiheadAttention.forward = networks.network_MultiheadAttention_forward
+torch.nn.MultiheadAttention._load_from_state_dict = networks.network_MultiheadAttention_load_state_dict
 
-script_callbacks.on_model_loaded(lora.assign_lora_names_to_compvis_modules)
+script_callbacks.on_model_loaded(networks.assign_network_names_to_compvis_modules)
 script_callbacks.on_script_unloaded(unload)
 script_callbacks.on_before_ui(before_ui)
-script_callbacks.on_infotext_pasted(lora.infotext_pasted)
+script_callbacks.on_infotext_pasted(networks.infotext_pasted)
 
 
 shared.options_templates.update(shared.options_section(('extra_networks', "Extra Networks"), {
-    "sd_lora": shared.OptionInfo("None", "Add Lora to prompt", gr.Dropdown, lambda: {"choices": ["None", *lora.available_loras]}, refresh=lora.list_available_loras),
+    "sd_lora": shared.OptionInfo("None", "Add network to prompt", gr.Dropdown, lambda: {"choices": ["None", *networks.available_networks]}, refresh=networks.list_available_networks),
     "lora_preferred_name": shared.OptionInfo("Alias from file", "When adding to prompt, refer to Lora by", gr.Radio, {"choices": ["Alias from file", "Filename"]}),
     "lora_add_hashes_to_infotext": shared.OptionInfo(True, "Add Lora hashes to infotext"),
 }))
 
 
 shared.options_templates.update(shared.options_section(('compatibility', "Compatibility"), {
-    "lora_functional": shared.OptionInfo(False, "Lora: use old method that takes longer when you have multiple Loras active and produces same results as kohya-ss/sd-webui-additional-networks extension"),
+    "lora_functional": shared.OptionInfo(False, "Lora/Networks: use old method that takes longer when you have multiple Loras active and produces same results as kohya-ss/sd-webui-additional-networks extension"),
 }))
 
 
-def create_lora_json(obj: lora.LoraOnDisk):
+def create_lora_json(obj: network.NetworkOnDisk):
     return {
         "name": obj.name,
         "alias": obj.alias,
@@ -75,17 +76,17 @@ def create_lora_json(obj: lora.LoraOnDisk):
     }
 
 
-def api_loras(_: gr.Blocks, app: FastAPI):
+def api_networks(_: gr.Blocks, app: FastAPI):
     @app.get("/sdapi/v1/loras")
     async def get_loras():
-        return [create_lora_json(obj) for obj in lora.available_loras.values()]
+        return [create_lora_json(obj) for obj in networks.available_networks.values()]
 
     @app.post("/sdapi/v1/refresh-loras")
     async def refresh_loras():
-        return lora.list_available_loras()
+        return networks.list_available_networks()
 
 
-script_callbacks.on_app_started(api_loras)
+script_callbacks.on_app_started(api_networks)
 
 re_lora = re.compile("<lora:([^:]+):")
 
@@ -98,19 +99,19 @@ def infotext_pasted(infotext, d):
     hashes = [x.strip().split(':', 1) for x in hashes.split(",")]
     hashes = {x[0].strip().replace(",", ""): x[1].strip() for x in hashes}
 
-    def lora_replacement(m):
+    def network_replacement(m):
         alias = m.group(1)
         shorthash = hashes.get(alias)
         if shorthash is None:
             return m.group(0)
 
-        lora_on_disk = lora.available_lora_hash_lookup.get(shorthash)
-        if lora_on_disk is None:
+        network_on_disk = networks.available_network_hash_lookup.get(shorthash)
+        if network_on_disk is None:
             return m.group(0)
 
-        return f'<lora:{lora_on_disk.get_alias()}:'
+        return f'<lora:{network_on_disk.get_alias()}:'
 
-    d["Prompt"] = re.sub(re_lora, lora_replacement, d["Prompt"])
+    d["Prompt"] = re.sub(re_lora, network_replacement, d["Prompt"])
 
 
 script_callbacks.on_infotext_pasted(infotext_pasted)

extensions-builtin/Lora/ui_extra_networks_lora.py

 import os
-import lora
+import networks
 
 from modules import shared, ui_extra_networks
 from modules.ui_extra_networks import quote_js
@@ -11,10 +11,10 @@ class ExtraNetworksPageLora(ui_extra_networks.ExtraNetworksPage):
         super().__init__('Lora')
 
     def refresh(self):
-        lora.list_available_loras()
+        networks.list_available_networks()
 
     def create_item(self, name, index=None):
-        lora_on_disk = lora.available_loras.get(name)
+        lora_on_disk = networks.available_networks.get(name)
 
         path, ext = os.path.splitext(lora_on_disk.filename)
 
@@ -43,7 +43,7 @@ class ExtraNetworksPageLora(ui_extra_networks.ExtraNetworksPage):
         return item
 
     def list_items(self):
-        for index, name in enumerate(lora.available_loras):
+        for index, name in enumerate(networks.available_networks):
             item = self.create_item(name, index)
             yield item