Merge pull request #3496 from vladmandic/lora-extract

Lora extract

extensions-builtin/Lora/lora_extract.py

@@ -0,0 +1,272 @@
+import os
+import time
+import json
+import datetime
+import torch
+from safetensors.torch import save_file
+import gradio as gr
+from modules import shared, devices
+from modules.ui_common import create_refresh_button
+from modules.call_queue import wrap_gradio_gpu_call
+
+
+class SVDHandler:
+    def __init__(self, maxrank=0, rank_ratio=1):
+        self.network_name: str = None
+        self.U: torch.Tensor = None
+        self.S: torch.Tensor = None
+        self.Vh: torch.Tensor = None
+        self.maxrank: int = maxrank
+        self.rank_ratio: float = rank_ratio
+        self.rank: int = 0
+        self.out_size: int = None
+        self.in_size: int = None
+        self.kernel_size: tuple[int, int] = None
+        self.conv2d: bool = False
+
+    def decompose(self, weight, backupweight):
+        self.conv2d = len(weight.size()) == 4
+        self.kernel_size = None if not self.conv2d else weight.size()[2:4]
+        self.out_size, self.in_size = weight.size()[0:2]
+        diffweight = weight.clone().to(devices.device)
+        diffweight -= backupweight.to(devices.device)
+        if self.conv2d:
+            if self.conv2d and self.kernel_size != (1, 1):
+                diffweight = diffweight.flatten(start_dim=1)
+            else:
+                diffweight = diffweight.squeeze()
+        self.U, self.S, self.Vh = torch.svd_lowrank(diffweight.to(device=devices.device, dtype=torch.float), self.maxrank, 2)
+        # del diffweight
+        self.U = self.U.to(device=devices.cpu, dtype=torch.bfloat16)
+        self.S = self.S.to(device=devices.cpu, dtype=torch.bfloat16)
+        self.Vh = self.Vh.t().to(device=devices.cpu, dtype=torch.bfloat16)  # svd_lowrank outputs a transposed matrix
+
+    def findrank(self):
+        if self.rank_ratio < 1:
+            S_squared = self.S.pow(2)
+            S_fro_sq = float(torch.sum(S_squared))
+            sum_S_squared = torch.cumsum(S_squared, dim=0) / S_fro_sq
+            index = int(torch.searchsorted(sum_S_squared, self.rank_ratio ** 2)) + 1
+            index = max(1, min(index, len(self.S) - 1))
+            self.rank = index
+            if self.maxrank > 0:
+                self.rank = min(self.rank, self.maxrank)
+        else:
+            self.rank = min(self.in_size, self.out_size, self.maxrank)
+
+    def makeweights(self):
+        self.findrank()
+        up = self.U[:, :self.rank] @ torch.diag(self.S[:self.rank])
+        down = self.Vh[:self.rank, :]
+        if self.conv2d and self.kernel_size is not None:
+            up = up.reshape(self.out_size, self.rank, 1, 1)
+            down = down.reshape(self.rank, self.in_size, self.kernel_size[0], self.kernel_size[1]) # pylint: disable=unsubscriptable-object
+        return_dict = {f'{self.network_name}.lora_up.weight': up.contiguous(),
+                       f'{self.network_name}.lora_down.weight': down.contiguous(),
+                       f'{self.network_name}.alpha': torch.tensor(down.shape[0]),
+                       }
+        return return_dict
+
+
+def loaded_lora():
+    if not shared.sd_loaded:
+        return ""
+    loaded = set()
+    if hasattr(shared.sd_model, 'unet'):
+        for _name, module in shared.sd_model.unet.named_modules():
+            current = getattr(module, "network_current_names", None)
+            if current is not None:
+                current = [item[0] for item in current]
+                loaded.update(current)
+    return list(loaded)
+
+
+def loaded_lora_str():
+    return ", ".join(loaded_lora())
+
+
+def make_meta(fn, maxrank, rank_ratio):
+    meta = {
+        "model_spec.sai_model_spec": "1.0.0",
+        "model_spec.title": os.path.splitext(os.path.basename(fn))[0],
+        "model_spec.author": "SD.Next",
+        "model_spec.implementation": "https://github.com/vladmandic/automatic",
+        "model_spec.date": datetime.datetime.now().astimezone().replace(microsecond=0).isoformat(),
+        "model_spec.base_model": shared.opts.sd_model_checkpoint,
+        "model_spec.dtype": str(devices.dtype),
+        "model_spec.base_lora": json.dumps(loaded_lora()),
+        "model_spec.config": f"maxrank={maxrank} rank_ratio={rank_ratio}",
+    }
+    if shared.sd_model_type == "sdxl":
+        meta["model_spec.architecture"] = "stable-diffusion-xl-v1-base/lora" # sai standard
+        meta["ss_base_model_version"] = "sdxl_base_v1-0" # kohya standard
+    elif shared.sd_model_type == "sd":
+        meta["model_spec.architecture"] = "stable-diffusion-v1/lora"
+        meta["ss_base_model_version"] = "sd_v1"
+    elif shared.sd_model_type == "f1":
+        meta["model_spec.architecture"] = "flux-1-dev/lora"
+        meta["ss_base_model_version"] = "flux1"
+    elif shared.sd_model_type == "sc":
+        meta["model_spec.architecture"] = "stable-cascade-v1-prior/lora"
+    return meta
+
+
+def make_lora(fn, maxrank, auto_rank, rank_ratio, modules, overwrite):
+    if not shared.sd_loaded or not shared.native:
+        msg = "LoRA extract: model not loaded"
+        shared.log.warning(msg)
+        yield msg
+        return
+    if loaded_lora() == "":
+        msg = "LoRA extract: no LoRA detected"
+        shared.log.warning(msg)
+        yield msg
+        return
+    if not fn:
+        msg = "LoRA extract: target filename required"
+        shared.log.warning(msg)
+        yield msg
+        return
+    t0 = time.time()
+    maxrank = int(maxrank)
+    rank_ratio = 1 if not auto_rank else rank_ratio
+    shared.log.debug(f'LoRA extract: modules={modules} maxrank={maxrank} auto={auto_rank} ratio={rank_ratio} fn="{fn}"')
+    shared.state.begin('LoRA extract')
+
+    # bar_format='Progress {rate_fmt}{postfix} {bar} {percentage:3.0f}% {n_fmt}/{total_fmt} {elapsed} {remaining} ' + '\x1b[38;5;71m' + desc, ncols=80, colour='#327fba'
+    from rich.progress import Progress, TextColumn, BarColumn, TaskProgressColumn, TimeRemainingColumn, TimeElapsedColumn
+    with Progress(TextColumn('[cyan]LoRA extract'), BarColumn(), TaskProgressColumn(), TimeRemainingColumn(), TimeElapsedColumn(), TextColumn('[cyan]{task.description}'), console=shared.console) as progress:
+
+        if 'te' in modules and getattr(shared.sd_model, 'text_encoder', None) is not None:
+            modules = shared.sd_model.text_encoder.named_modules()
+            task = progress.add_task(description="te1 decompose", total=len(list(modules)))
+            for name, module in shared.sd_model.text_encoder.named_modules():
+                progress.update(task, advance=1)
+                weights_backup = getattr(module, "network_weights_backup", None)
+                if weights_backup is None or getattr(module, "network_current_names", None) is None:
+                    continue
+                prefix = "lora_te1_" if hasattr(shared.sd_model, 'text_encoder_2') else "lora_te_"
+                module.svdhandler = SVDHandler(maxrank, rank_ratio)
+                module.svdhandler.network_name = prefix + name.replace(".", "_")
+                with devices.inference_context():
+                    module.svdhandler.decompose(module.weight, weights_backup)
+            progress.remove_task(task)
+        t1 = time.time()
+
+        if 'te' in modules and getattr(shared.sd_model, 'text_encoder_2', None) is not None:
+            modules = shared.sd_model.text_encoder_2.named_modules()
+            task = progress.add_task(description="te2 decompose", total=len(list(modules)))
+            for name, module in shared.sd_model.text_encoder_2.named_modules():
+                progress.update(task, advance=1)
+                weights_backup = getattr(module, "network_weights_backup", None)
+                if weights_backup is None or getattr(module, "network_current_names", None) is None:
+                    continue
+                module.svdhandler = SVDHandler(maxrank, rank_ratio)
+                module.svdhandler.network_name = "lora_te2_" + name.replace(".", "_")
+                with devices.inference_context():
+                    module.svdhandler.decompose(module.weight, weights_backup)
+            progress.remove_task(task)
+        t2 = time.time()
+
+        if 'unet' in modules and getattr(shared.sd_model, 'unet', None) is not None:
+            modules = shared.sd_model.unet.named_modules()
+            task = progress.add_task(description="unet decompose", total=len(list(modules)))
+            for name, module in shared.sd_model.unet.named_modules():
+                progress.update(task, advance=1)
+                weights_backup = getattr(module, "network_weights_backup", None)
+                if weights_backup is None or getattr(module, "network_current_names", None) is None:
+                    continue
+                module.svdhandler = SVDHandler(maxrank, rank_ratio)
+                module.svdhandler.network_name = "lora_unet_" + name.replace(".", "_")
+                with devices.inference_context():
+                    module.svdhandler.decompose(module.weight, weights_backup)
+            progress.remove_task(task)
+        t3 = time.time()
+
+        # TODO: Handle quant for Flux
+        # if 'te' in modules and getattr(shared.sd_model, 'transformer', None) is not None:
+        #     for name, module in shared.sd_model.transformer.named_modules():
+        #         if "norm" in name and "linear" not in name:
+        #             continue
+        #         weights_backup = getattr(module, "network_weights_backup", None)
+        #         if weights_backup is None:
+        #             continue
+        #         module.svdhandler = SVDHandler()
+        #         module.svdhandler.network_name = "lora_transformer_" + name.replace(".", "_")
+        #         module.svdhandler.decompose(module.weight, weights_backup)
+        #         module.svdhandler.findrank(rank, rank_ratio)
+
+        lora_state_dict = {}
+        for sub in ['text_encoder', 'text_encoder_2', 'unet', 'transformer']:
+            submodel = getattr(shared.sd_model, sub, None)
+            if submodel is not None:
+                modules = submodel.named_modules()
+                task = progress.add_task(description=f"{sub} exctract", total=len(list(modules)))
+                for _name, module in submodel.named_modules():
+                    progress.update(task, advance=1)
+                    if not hasattr(module, "svdhandler"):
+                        continue
+                    lora_state_dict.update(module.svdhandler.makeweights())
+                    del module.svdhandler
+                progress.remove_task(task)
+        t4 = time.time()
+
+    if not os.path.isabs(fn):
+        fn = os.path.join(shared.cmd_opts.lora_dir, fn)
+    if not fn.endswith('.safetensors'):
+        fn += '.safetensors'
+    if os.path.exists(fn):
+        if overwrite:
+            os.remove(fn)
+        else:
+            msg = f'LoRA extract: fn="{fn}" file exists'
+            shared.log.warning(msg)
+            yield msg
+            return
+
+    shared.state.end()
+    meta = make_meta(fn, maxrank, rank_ratio)
+    shared.log.debug(f'LoRA metadata: {meta}')
+    try:
+        save_file(tensors=lora_state_dict, metadata=meta, filename=fn)
+    except Exception as e:
+        msg = f'LoRA extract error: fn="{fn}" {e}'
+        shared.log.error(msg)
+        yield msg
+        return
+    t5 = time.time()
+    shared.log.debug(f'LoRA extract: time={t5-t0:.2f} te1={t1-t0:.2f} te2={t2-t1:.2f} unet={t3-t2:.2f} save={t5-t4:.2f}')
+    keys = list(lora_state_dict.keys())
+    msg = f'LoRA extract: fn="{fn}" keys={len(keys)}'
+    shared.log.info(msg)
+    yield msg
+
+
+def create_ui():
+    def gr_show(visible=True):
+        return {"visible": visible, "__type__": "update"}
+
+    with gr.Tab(label="Extract LoRA"):
+        with gr.Row():
+            loaded = gr.Textbox(placeholder="Press refresh to query loaded LoRA", label="Loaded LoRA", interactive=False)
+            create_refresh_button(loaded, lambda: None, lambda: {'value': loaded_lora_str()}, "testid")
+        with gr.Group():
+            with gr.Row():
+                modules = gr.CheckboxGroup(label="Modules to extract", value=['unet'], choices=['te', 'unet'])
+            with gr.Row():
+                auto_rank = gr.Checkbox(value=False, label="Automatically determine rank")
+                rank_ratio = gr.Slider(label="Autorank ratio", value=1, minimum=0, maximum=1, step=0.05, visible=False)
+                rank = gr.Slider(label="Maximum rank", value=32, minimum=1, maximum=256)
+        with gr.Row():
+            filename = gr.Textbox(label="LoRA target filename")
+            overwrite = gr.Checkbox(value=False, label="Overwrite existing file")
+        with gr.Row():
+            extract = gr.Button(value="Extract LoRA", variant='primary')
+            status = gr.HTML(value="", show_label=False)
+
+        auto_rank.change(fn=lambda x: gr_show(x), inputs=[auto_rank], outputs=[rank_ratio])
+        extract.click(
+            fn=wrap_gradio_gpu_call(make_lora, extra_outputs=[]),
+            inputs=[filename, rank, auto_rank, rank_ratio, modules, overwrite],
+            outputs=[status]
+        )

extensions-builtin/Lora/networks.py

@@ -285,6 +285,8 @@ def network_restore_weights_from_backup(self: Union[torch.nn.Conv2d, torch.nn.Li
     weights_backup = getattr(self, "network_weights_backup", None)
     bias_backup = getattr(self, "network_bias_backup", None)
     if weights_backup is None and bias_backup is None:
+        t1 = time.time()
+        timer['restore'] += t1 - t0
         return
     # if debug:
     #     shared.log.debug('LoRA restore weights')
@@ -319,18 +321,7 @@ def network_restore_weights_from_backup(self: Union[torch.nn.Conv2d, torch.nn.Li
     timer['restore'] += t1 - t0
 
 
-def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn.GroupNorm, torch.nn.LayerNorm, torch.nn.MultiheadAttention, diffusers.models.lora.LoRACompatibleLinear, diffusers.models.lora.LoRACompatibleConv]):
-    """
-    Applies the currently selected set of networks to the weights of torch layer self.
-    If weights already have this particular set of networks applied, does nothing.
-    If not, restores orginal weights from backup and alters weights according to networks.
-    """
-    network_layer_name = getattr(self, 'network_layer_name', None)
-    if network_layer_name is None:
-        return
-    t0 = time.time()
-    current_names = getattr(self, "network_current_names", ())
-    wanted_names = tuple((x.name, x.te_multiplier, x.unet_multiplier, x.dyn_dim) for x in loaded_networks)
+def maybe_backup_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn.GroupNorm, torch.nn.LayerNorm, torch.nn.MultiheadAttention, diffusers.models.lora.LoRACompatibleLinear, diffusers.models.lora.LoRACompatibleConv], wanted_names, current_names):
     weights_backup = getattr(self, "network_weights_backup", None)
     if weights_backup is None and wanted_names != (): # pylint: disable=C1803
         if current_names != ():
@@ -360,6 +351,21 @@ def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn
             bias_backup = None
         self.network_bias_backup = bias_backup
 
+
+def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn.GroupNorm, torch.nn.LayerNorm, torch.nn.MultiheadAttention, diffusers.models.lora.LoRACompatibleLinear, diffusers.models.lora.LoRACompatibleConv]):
+    """
+    Applies the currently selected set of networks to the weights of torch layer self.
+    If weights already have this particular set of networks applied, does nothing.
+    If not, restores orginal weights from backup and alters weights according to networks.
+    """
+    network_layer_name = getattr(self, 'network_layer_name', None)
+    if network_layer_name is None:
+        return
+    t0 = time.time()
+    current_names = getattr(self, "network_current_names", ())
+    wanted_names = tuple((x.name, x.te_multiplier, x.unet_multiplier, x.dyn_dim) for x in loaded_networks)
+    if any([net.modules.get(network_layer_name, None) for net in loaded_networks]):
+        maybe_backup_weights(self, wanted_names, current_names)
     if current_names != wanted_names:
         network_restore_weights_from_backup(self)
         for net in loaded_networks:

extensions-builtin/Lora/scripts/lora_script.py

@@ -1,6 +1,7 @@
 import re
 import networks
 import lora # pylint: disable=unused-import
+from lora_extract import create_ui
 from network import NetworkOnDisk
 from ui_extra_networks_lora import ExtraNetworksPageLora
 from extra_networks_lora import ExtraNetworkLora
@@ -14,6 +15,7 @@ def before_ui():
     ui_extra_networks.register_page(ExtraNetworksPageLora())
     networks.extra_network_lora = ExtraNetworkLora()
     extra_networks.register_extra_network(networks.extra_network_lora)
+    ui_models.extra_ui.append(create_ui)
 
 
 def create_lora_json(obj: NetworkOnDisk):

modules/devices.py

@@ -508,12 +508,12 @@ def test_for_nans(x, where):
     raise NansException(message)
 
 
-def normalize_device(device):
-    if torch.device(device).type in {"cpu", "mps", "meta"}:
-        return torch.device(device)
-    if torch.device(device).index is None:
-        return torch.device(str(device), index=0)
-    return torch.device(device)
+def normalize_device(dev):
+    if torch.device(dev).type in {"cpu", "mps", "meta"}:
+        return torch.device(dev)
+    if torch.device(dev).index is None:
+        return torch.device(str(dev), index=0)
+    return torch.device(dev)
 
 
 def same_device(d1, d2):

modules/ui_sections.py

@@ -270,7 +270,7 @@ def create_sampler_options(tabname):
             sampler_options = gr.CheckboxGroup(label='Sampler options', elem_id=f"{tabname}_sampler_options", choices=options, value=values, type='value')
         with gr.Row(elem_classes=['flex-break']):
             shared.opts.data['schedulers_sigma'] = shared.opts.data.get('schedulers_sigma', 'default')
-            sampler_algo = gr.Radio(label='Sigma algorithm', elem_id=f"{tabname}_sigma_algo", choices=['default', 'karras', 'exponential', 'polyexponential'], value=shared.opts.data['schedulers_sigma'], type='value')
+            sampler_algo = gr.Radio(label='Sigma algorithm', elem_id=f"{tabname}_sigma_algo", choices=['default', 'karras', 'exponential', 'polyexponential'], value=shared.opts.data.schedulers_sigma, type='value')
         sampler_options.change(fn=set_sampler_original_options, inputs=[sampler_options, sampler_algo], outputs=[])
         sampler_algo.change(fn=set_sampler_original_options, inputs=[sampler_options, sampler_algo], outputs=[])