From 0c3feb202c5714abd50d879c1db2cd9a71ce93e3 Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Tue, 10 Jan 2023 14:08:29 +0300
Subject: [PATCH] disable torch weight initialization and CLIP
 downloading/reading checkpoint to speedup creating sd model from config

---
 modules/sd_disable_initialization.py | 44 ++++++++++++++++++++++++++++
 modules/sd_models.py                 |  5 ++--
 2 files changed, 47 insertions(+), 2 deletions(-)
 create mode 100644 modules/sd_disable_initialization.py

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)