disable torch weight initialization and CLIP downloading/reading checkpoint to speedup creating sd model from config

2 files changed, 47 insertions, 2 deletions

diff --git a/modules/sd_disable_initialization.py b/modules/sd_disable_initialization.py
new file mode 100644
index 00000000..c9a3b5e4
--- /dev/null
+++ b/modules/sd_disable_initialization.py
@@ -0,0 +1,44 @@
+import ldm.modules.encoders.modules

+import open_clip

+import torch

+

+

+class DisableInitialization:

+    """

+    When an object of this class enters a `with` block, it starts preventing torch's layer initialization

+    functions from working, and changes CLIP and OpenCLIP to not download model weights. When it leaves,

+    reverts everything to how it was.

+

+    Use like this:

+    ```

+    with DisableInitialization():

+        do_things()

+    ```

+    """

+

+    def __enter__(self):

+        def do_nothing(*args, **kwargs):

+            pass

+

+        def create_model_and_transforms_without_pretrained(*args, pretrained=None, **kwargs):

+            return self.create_model_and_transforms(*args, pretrained=None, **kwargs)

+

+        def CLIPTextModel_from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs):

+            return self.CLIPTextModel_from_pretrained(None, *model_args, config=pretrained_model_name_or_path, state_dict={}, **kwargs)

+

+        self.init_kaiming_uniform = torch.nn.init.kaiming_uniform_

+        self.init_no_grad_normal = torch.nn.init._no_grad_normal_

+        self.create_model_and_transforms = open_clip.create_model_and_transforms

+        self.CLIPTextModel_from_pretrained = ldm.modules.encoders.modules.CLIPTextModel.from_pretrained

+

+        torch.nn.init.kaiming_uniform_ = do_nothing

+        torch.nn.init._no_grad_normal_ = do_nothing

+        open_clip.create_model_and_transforms = create_model_and_transforms_without_pretrained

+        ldm.modules.encoders.modules.CLIPTextModel.from_pretrained = CLIPTextModel_from_pretrained

+

+    def __exit__(self, exc_type, exc_val, exc_tb):

+        torch.nn.init.kaiming_uniform_ = self.init_kaiming_uniform

+        torch.nn.init._no_grad_normal_ = self.init_no_grad_normal

+        open_clip.create_model_and_transforms = self.create_model_and_transforms

+        ldm.modules.encoders.modules.CLIPTextModel.from_pretrained = self.CLIPTextModel_from_pretrained

+

diff --git a/modules/sd_models.py b/modules/sd_models.py
index 0a6d55ca..ee241032 100644
--- a/modules/sd_models.py
+++ b/modules/sd_models.py
@@ -13,7 +13,7 @@ import ldm.modules.midas as midas
 

 from ldm.util import instantiate_from_config

 

-from modules import shared, modelloader, devices, script_callbacks, sd_vae

+from modules import shared, modelloader, devices, script_callbacks, sd_vae, sd_disable_initialization

 from modules.paths import models_path

 from modules.sd_hijack_inpainting import do_inpainting_hijack, should_hijack_inpainting

 

@@ -319,7 +319,8 @@ def load_model(checkpoint_info=None):
     if shared.cmd_opts.no_half:

         sd_config.model.params.unet_config.params.use_fp16 = False

 

-    sd_model = instantiate_from_config(sd_config.model)

+    with sd_disable_initialization.DisableInitialization():

+        sd_model = instantiate_from_config(sd_config.model)

 

     load_model_weights(sd_model, checkpoint_info)