Reload VAE without reloading sd checkpoint

Author: R-N

modules/sd_models.py

@@ -159,15 +159,13 @@ def get_state_dict_from_checkpoint(pl_sd):
     return pl_sd
 
 
-vae_ignore_keys = {"model_ema.decay", "model_ema.num_updates"}
-
 def load_model_weights(model, checkpoint_info, vae_file="auto"):
     checkpoint_file = checkpoint_info.filename
     sd_model_hash = checkpoint_info.hash
 
     vae_file = sd_vae.resolve_vae(checkpoint_file, vae_file=vae_file)
 
-    checkpoint_key = (checkpoint_info, vae_file)
+    checkpoint_key = checkpoint_info
 
     if checkpoint_key not in checkpoints_loaded:
         print(f"Loading weights [{sd_model_hash}] from {checkpoint_file}")
@@ -190,13 +188,12 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"):
         devices.dtype = torch.float32 if shared.cmd_opts.no_half else torch.float16
         devices.dtype_vae = torch.float32 if shared.cmd_opts.no_half or shared.cmd_opts.no_half_vae else torch.float16
 
-        sd_vae.load_vae(model, vae_file)
-        model.first_stage_model.to(devices.dtype_vae)
-
         if shared.opts.sd_checkpoint_cache > 0:
+            # if PR #4035 were to get merged, restore base VAE first before caching
             checkpoints_loaded[checkpoint_key] = model.state_dict().copy()
             while len(checkpoints_loaded) > shared.opts.sd_checkpoint_cache:
                 checkpoints_loaded.popitem(last=False)  # LRU
+
     else:
         vae_name = sd_vae.get_filename(vae_file)
         print(f"Loading weights [{sd_model_hash}] with {vae_name} VAE from cache")
@@ -207,6 +204,8 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"):
     model.sd_model_checkpoint = checkpoint_file
     model.sd_checkpoint_info = checkpoint_info
 
+    sd_vae.load_vae(model, vae_file)
+
 
 def load_model(checkpoint_info=None):
     from modules import lowvram, sd_hijack
@@ -254,14 +253,14 @@ def load_model(checkpoint_info=None):
     return sd_model
 
 
-def reload_model_weights(sd_model=None, info=None, force=False):
+def reload_model_weights(sd_model=None, info=None):
     from modules import lowvram, devices, sd_hijack
     checkpoint_info = info or select_checkpoint()
 
     if not sd_model:
         sd_model = shared.sd_model
 
-    if sd_model.sd_model_checkpoint == checkpoint_info.filename and not force:
+    if sd_model.sd_model_checkpoint == checkpoint_info.filename:
         return
 
     if sd_model.sd_checkpoint_info.config != checkpoint_info.config or should_hijack_inpainting(checkpoint_info) != should_hijack_inpainting(sd_model.sd_checkpoint_info):

modules/sd_vae.py

@@ -1,26 +1,65 @@
 import torch
 import os
 from collections import namedtuple
-from modules import shared, devices
+from modules import shared, devices, script_callbacks
 from modules.paths import models_path
 import glob
 
+
 model_dir = "Stable-diffusion"
 model_path = os.path.abspath(os.path.join(models_path, model_dir))
 vae_dir = "VAE"
 vae_path = os.path.abspath(os.path.join(models_path, vae_dir))
 
+
 vae_ignore_keys = {"model_ema.decay", "model_ema.num_updates"}
+
+
 default_vae_dict = {"auto": "auto", "None": "None"}
 default_vae_list = ["auto", "None"]
+
+
 default_vae_values = [default_vae_dict[x] for x in default_vae_list]
 vae_dict = dict(default_vae_dict)
 vae_list = list(default_vae_list)
 first_load = True
 
+
+base_vae = None
+loaded_vae_file = None
+checkpoint_info = None
+
+
+def get_base_vae(model):
+    if base_vae is not None and checkpoint_info == model.sd_checkpoint_info and model:
+        return base_vae
+    return None
+
+
+def store_base_vae(model):
+    global base_vae, checkpoint_info
+    if checkpoint_info != model.sd_checkpoint_info:
+        base_vae = model.first_stage_model.state_dict().copy()
+        checkpoint_info = model.sd_checkpoint_info
+
+
+def delete_base_vae():
+    global base_vae, checkpoint_info
+    base_vae = None
+    checkpoint_info = None
+
+
+def restore_base_vae(model):
+    global base_vae, checkpoint_info
+    if base_vae is not None and checkpoint_info == model.sd_checkpoint_info:
+        load_vae_dict(model, base_vae)
+    delete_base_vae()
+
+
 def get_filename(filepath):
     return os.path.splitext(os.path.basename(filepath))[0]
 
+
 def refresh_vae_list(vae_path=vae_path, model_path=model_path):
     global vae_dict, vae_list
     res = {}
@@ -43,6 +82,7 @@ def refresh_vae_list(vae_path=vae_path, model_path=model_path):
     vae_dict.update(res)
     return vae_list
 
+
 def resolve_vae(checkpoint_file, vae_file="auto"):
     global first_load, vae_dict, vae_list
     # save_settings = False
@@ -96,24 +136,26 @@ def resolve_vae(checkpoint_file, vae_file="auto"):
 
     return vae_file
 
-def load_vae(model, vae_file):
-    global first_load, vae_dict, vae_list
+
+def load_vae(model, vae_file=None):
+    global first_load, vae_dict, vae_list, loaded_vae_file
     # save_settings = False
 
     if vae_file:
         print(f"Loading VAE weights from: {vae_file}")
         vae_ckpt = torch.load(vae_file, map_location=shared.weight_load_location)
         vae_dict_1 = {k: v for k, v in vae_ckpt["state_dict"].items() if k[0:4] != "loss" and k not in vae_ignore_keys}
-        model.first_stage_model.load_state_dict(vae_dict_1)
+        load_vae_dict(model, vae_dict_1)
 
-    # If vae used is not in dict, update it
-    # It will be removed on refresh though
-    if vae_file is not None:
+        # If vae used is not in dict, update it
+        # It will be removed on refresh though
         vae_opt = get_filename(vae_file)
         if vae_opt not in vae_dict:
             vae_dict[vae_opt] = vae_file
             vae_list.append(vae_opt)
 
+    loaded_vae_file = vae_file
+
     """
     # Save current VAE to VAE settings, maybe? will it work?
     if save_settings:
@@ -124,4 +166,45 @@ def load_vae(model, vae_file):
     """
 
     first_load = False
+
+
+# don't call this from outside
+def load_vae_dict(model, vae_dict_1=None):
+    if vae_dict_1:
+        store_base_vae(model)
+        model.first_stage_model.load_state_dict(vae_dict_1)
+    else:
+        restore_base_vae()
     model.first_stage_model.to(devices.dtype_vae)
+
+
+def reload_vae_weights(sd_model=None, vae_file="auto"):
+    from modules import lowvram, devices, sd_hijack
+
+    if not sd_model:
+        sd_model = shared.sd_model
+
+    checkpoint_info = sd_model.sd_checkpoint_info
+    checkpoint_file = checkpoint_info.filename
+    vae_file = resolve_vae(checkpoint_file, vae_file=vae_file)
+
+    if loaded_vae_file == vae_file:
+        return
+
+    if shared.cmd_opts.lowvram or shared.cmd_opts.medvram:
+        lowvram.send_everything_to_cpu()
+    else:
+        sd_model.to(devices.cpu)
+
+    sd_hijack.model_hijack.undo_hijack(sd_model)
+
+    load_vae(sd_model, vae_file)
+
+    sd_hijack.model_hijack.hijack(sd_model)
+    script_callbacks.model_loaded_callback(sd_model)
+
+    if not shared.cmd_opts.lowvram and not shared.cmd_opts.medvram:
+        sd_model.to(devices.device)
+
+    print(f"VAE Weights loaded.")
+    return sd_model

webui.py

@@ -81,9 +81,7 @@ def initialize():
     modules.sd_vae.refresh_vae_list()
     modules.sd_models.load_model()
     shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: modules.sd_models.reload_model_weights()))
-    # I don't know what needs to be done to only reload VAE, with all those hijacks callbacks, and lowvram, 
-    # so for now this reloads the whole model too
-    shared.opts.onchange("sd_vae", wrap_queued_call(lambda: modules.sd_models.reload_model_weights(force=True)), call=False)
+    shared.opts.onchange("sd_vae", wrap_queued_call(lambda: modules.sd_vae.reload_vae_weights()), call=False)
     shared.opts.onchange("sd_hypernetwork", wrap_queued_call(lambda: modules.hypernetworks.hypernetwork.load_hypernetwork(shared.opts.sd_hypernetwork)))
     shared.opts.onchange("sd_hypernetwork_strength", modules.hypernetworks.hypernetwork.apply_strength)