Skip weight initialization when resizing text encoder token embedding…

…s to accomodate new TI embeddings. This saves time.

invokeai/backend/model_management/lora.py

@@ -13,6 +13,7 @@
 from transformers import CLIPTextModel, CLIPTokenizer
 
 from invokeai.app.shared.models import FreeUConfig
+from invokeai.backend.model_management.model_load_optimizations import skip_torch_weight_init
 
 from .models.lora import LoRAModel
 
@@ -211,8 +212,12 @@ def _get_ti_embedding(model_embeddings, ti):
                 for i in range(ti_embedding.shape[0]):
                     new_tokens_added += ti_tokenizer.add_tokens(_get_trigger(ti_name, i))
 
-            # modify text_encoder
-            text_encoder.resize_token_embeddings(init_tokens_count + new_tokens_added, pad_to_multiple_of)
+            # Modify text_encoder.
+            # resize_token_embeddings(...) constructs a new torch.nn.Embedding internally. Initializing the weights of
+            # this embedding is slow and unnecessary, so we wrap this step in skip_torch_weight_init() to save some
+            # time.
+            with skip_torch_weight_init():
+                text_encoder.resize_token_embeddings(init_tokens_count + new_tokens_added, pad_to_multiple_of)
             model_embeddings = text_encoder.get_input_embeddings()
 
             for ti_name, ti in ti_list: