add an option to unload models during hypernetwork training to save VRAM

modules/hypernetworks/hypernetwork.py

@@ -175,6 +175,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, data_root, log_directory,
     filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork_name}.pt')
 
     log_directory = os.path.join(log_directory, datetime.datetime.now().strftime("%Y-%m-%d"), hypernetwork_name)
+    unload = shared.opts.unload_models_when_training
 
     if save_hypernetwork_every > 0:
         hypernetwork_dir = os.path.join(log_directory, "hypernetworks")
@@ -188,11 +189,13 @@ def train_hypernetwork(hypernetwork_name, learn_rate, data_root, log_directory,
     else:
         images_dir = None
 
-    cond_model = shared.sd_model.cond_stage_model
-
     shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..."
     with torch.autocast("cuda"):
-        ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=512, height=512, repeats=1, placeholder_token=hypernetwork_name, model=shared.sd_model, device=devices.device, template_file=template_file)
+        ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=512, height=512, repeats=1, placeholder_token=hypernetwork_name, model=shared.sd_model, device=devices.device, template_file=template_file, include_cond=True)
+
+    if unload:
+        shared.sd_model.cond_stage_model.to(devices.cpu)
+        shared.sd_model.first_stage_model.to(devices.cpu)
 
     hypernetwork = shared.loaded_hypernetwork
     weights = hypernetwork.weights()
@@ -211,7 +214,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, data_root, log_directory,
         return hypernetwork, filename
 
     pbar = tqdm.tqdm(enumerate(ds), total=steps - ititial_step)
-    for i, (x, text) in pbar:
+    for i, (x, text, cond) in pbar:
         hypernetwork.step = i + ititial_step
 
         if hypernetwork.step > steps:
@@ -221,11 +224,11 @@ def train_hypernetwork(hypernetwork_name, learn_rate, data_root, log_directory,
             break
 
         with torch.autocast("cuda"):
-            c = cond_model([text])
-
+            cond = cond.to(devices.device)
             x = x.to(devices.device)
-            loss = shared.sd_model(x.unsqueeze(0), c)[0]
+            loss = shared.sd_model(x.unsqueeze(0), cond)[0]
             del x
+            del cond
 
             losses[hypernetwork.step % losses.shape[0]] = loss.item()
 
@@ -244,6 +247,10 @@ def train_hypernetwork(hypernetwork_name, learn_rate, data_root, log_directory,
 
             preview_text = text if preview_image_prompt == "" else preview_image_prompt
 
+            optimizer.zero_grad()
+            shared.sd_model.cond_stage_model.to(devices.device)
+            shared.sd_model.first_stage_model.to(devices.device)
+
             p = processing.StableDiffusionProcessingTxt2Img(
                 sd_model=shared.sd_model,
                 prompt=preview_text,
@@ -255,6 +262,10 @@ def train_hypernetwork(hypernetwork_name, learn_rate, data_root, log_directory,
             processed = processing.process_images(p)
             image = processed.images[0]
 
+            if unload:
+                shared.sd_model.cond_stage_model.to(devices.cpu)
+                shared.sd_model.first_stage_model.to(devices.cpu)
+
             shared.state.current_image = image
             image.save(last_saved_image)
 

modules/hypernetworks/ui.py

@@ -5,7 +5,7 @@ import gradio as gr
 
 import modules.textual_inversion.textual_inversion
 import modules.textual_inversion.preprocess
-from modules import sd_hijack, shared
+from modules import sd_hijack, shared, devices
 from modules.hypernetworks import hypernetwork
 
 
@@ -41,5 +41,7 @@ Hypernetwork saved to {html.escape(filename)}
         raise
     finally:
         shared.loaded_hypernetwork = initial_hypernetwork
+        shared.sd_model.cond_stage_model.to(devices.device)
+        shared.sd_model.first_stage_model.to(devices.device)
         sd_hijack.apply_optimizations()
 

modules/shared.py

@@ -228,6 +228,10 @@ options_templates.update(options_section(('system', "System"), {
     "multiple_tqdm": OptionInfo(True, "Add a second progress bar to the console that shows progress for an entire job."),
 }))
 
+options_templates.update(options_section(('training', "Training"), {
+    "unload_models_when_training": OptionInfo(False, "Unload VAE and CLIP form VRAM when training"),
+}))
+
 options_templates.update(options_section(('sd', "Stable Diffusion"), {
     "sd_model_checkpoint": OptionInfo(None, "Stable Diffusion checkpoint", gr.Dropdown, lambda: {"choices": modules.sd_models.checkpoint_tiles()}, show_on_main_page=True),
     "sd_hypernetwork": OptionInfo("None", "Stable Diffusion finetune hypernetwork", gr.Dropdown, lambda: {"choices": ["None"] + [x for x in hypernetworks.keys()]}),

modules/textual_inversion/dataset.py

@@ -8,14 +8,14 @@ from torchvision import transforms
 
 import random
 import tqdm
-from modules import devices
+from modules import devices, shared
 import re
 
 re_tag = re.compile(r"[a-zA-Z][_\w\d()]+")
 
 
 class PersonalizedBase(Dataset):
-    def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, device=None, template_file=None):
+    def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, device=None, template_file=None, include_cond=False):
 
         self.placeholder_token = placeholder_token
 
@@ -32,6 +32,8 @@ class PersonalizedBase(Dataset):
 
         assert data_root, 'dataset directory not specified'
 
+        cond_model = shared.sd_model.cond_stage_model
+
         self.image_paths = [os.path.join(data_root, file_path) for file_path in os.listdir(data_root)]
         print("Preparing dataset...")
         for path in tqdm.tqdm(self.image_paths):
@@ -53,7 +55,13 @@ class PersonalizedBase(Dataset):
             init_latent = model.get_first_stage_encoding(model.encode_first_stage(torchdata.unsqueeze(dim=0))).squeeze()
             init_latent = init_latent.to(devices.cpu)
 
-            self.dataset.append((init_latent, filename_tokens))
+            if include_cond:
+                text = self.create_text(filename_tokens)
+                cond = cond_model([text]).to(devices.cpu)
+            else:
+                cond = None
+
+            self.dataset.append((init_latent, filename_tokens, cond))
 
         self.length = len(self.dataset) * repeats
 
@@ -64,6 +72,12 @@ class PersonalizedBase(Dataset):
     def shuffle(self):
         self.indexes = self.initial_indexes[torch.randperm(self.initial_indexes.shape[0])]
 
+    def create_text(self, filename_tokens):
+        text = random.choice(self.lines)
+        text = text.replace("[name]", self.placeholder_token)
+        text = text.replace("[filewords]", ' '.join(filename_tokens))
+        return text
+
     def __len__(self):
         return self.length
 
@@ -72,10 +86,7 @@ class PersonalizedBase(Dataset):
             self.shuffle()
 
         index = self.indexes[i % len(self.indexes)]
-        x, filename_tokens = self.dataset[index]
+        x, filename_tokens, cond = self.dataset[index]
 
-        text = random.choice(self.lines)
-        text = text.replace("[name]", self.placeholder_token)
-        text = text.replace("[filewords]", ' '.join(filename_tokens))
-
-        return x, text
+        text = self.create_text(filename_tokens)
+        return x, text, cond

modules/textual_inversion/textual_inversion.py

@@ -201,7 +201,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini
         return embedding, filename
 
     pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step)
-    for i, (x, text) in pbar:
+    for i, (x, text, _) in pbar:
         embedding.step = i + ititial_step
 
         if embedding.step > steps: