Merge branch 'dev' into Dev/DirectoryCacheing

CHANGELOG.md

@@ -102,6 +102,8 @@ And it also includes fixes for all reported issues so far
     - faster json parsing  
     - faster lora indexing  
     - lazy load optional imports  
+    - batch embedding load, thanks @midcoastal  
+      10x+ faster embeddings load for large number of embeddings, now works for 1000+ embeddings  
   - **extra networks**  
     - 4x faster civitai metadata and previews lookup  
     - better display and selection of tags & trigger words  

modules/patches.py

@@ -4,54 +4,40 @@ from typing import Optional
 
 def patch(key, obj, field, replacement, add_if_not_exists:bool = False):
     """Replaces a function in a module or a class.
-
     Also stores the original function in this module, possible to be retrieved via original(key, obj, field).
     If the function is already replaced by this caller (key), an exception is raised -- use undo() before that.
-
     Arguments:
         key: identifying information for who is doing the replacement. You can use __name__.
         obj: the module or the class
         field: name of the function as a string
         replacement: the new function
-
     Returns:
         the original function
     """
-
     patch_key = (obj, field)
     if patch_key in originals[key]:
         raise RuntimeError(f"patch for {field} is already applied")
-
     if not hasattr(obj, field) and not add_if_not_exists:
         raise AttributeError(f"type {type(obj)} '{type.__name__}' has no attribute '{field}'")
-
     original_func = getattr(obj, field, None)
     originals[key][patch_key] = original_func
-
     setattr(obj, field, replacement)
-
     return original_func
 
 
 def undo(key, obj, field):
     """Undoes the peplacement by the patch().
-
     If the function is not replaced, raises an exception.
-
     Arguments:
         key: identifying information for who is doing the replacement. You can use __name__.
         obj: the module or the class
         field: name of the function as a string
-
     Returns:
         Always None
     """
-
     patch_key = (obj, field)
-
     if patch_key not in originals[key]:
         raise RuntimeError(f"there is no patch for {field} to undo")
-
     original_func = originals[key].pop(patch_key)
     if original_func is None:
         delattr(obj, field)
@@ -62,7 +48,6 @@ def undo(key, obj, field):
 def original(key, obj, field):
     """Returns the original function for the patch created by the patch() function"""
     patch_key = (obj, field)
-
     return originals[key].get(patch_key, None)
 
 

modules/prompt_parser_diffusers.py

@@ -137,7 +137,7 @@ def prepare_embedding_providers(pipe, clip_skip):
         provider = EmbeddingsProvider(tokenizer=pipe.tokenizer, text_encoder=pipe.text_encoder, truncate=False,
                                       returned_embeddings_type=embedding_type, device=device)
         embeddings_providers.append(provider)
-    if hasattr(pipe, "tokenizer_2") and getattr(pipe, "text_encoder_2"):
+    if hasattr(pipe, "tokenizer_2") and hasattr(pipe, "text_encoder_2"):
         provider = EmbeddingsProvider(tokenizer=pipe.tokenizer_2, text_encoder=pipe.text_encoder_2, truncate=False,
                                       returned_embeddings_type=embedding_type, device=device)
         embeddings_providers.append(provider)

modules/textual_inversion/loaders.py

@@ -1,24 +1,19 @@
 from typing import TYPE_CHECKING, Dict, List, Optional, Union
-
 import torch
-from diffusers.loaders.textual_inversion import (
-    TextualInversionLoaderMixin,
-    load_textual_inversion_state_dicts,
-    logger,
-    nn,
-)
-
+from torch import nn
+from diffusers.loaders.textual_inversion import TextualInversionLoaderMixin, load_textual_inversion_state_dicts
 from modules import shared
 from modules.patches import patch_method
 
+
 if TYPE_CHECKING:
     from transformers import PreTrainedModel, PreTrainedTokenizer
-
 try:
     from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
 except Exception:
     pass
 
+
 @patch_method(TextualInversionLoaderMixin)
 def load_textual_inversion(
     self: TextualInversionLoaderMixin,
@@ -28,12 +23,10 @@ def load_textual_inversion(
     text_encoder: Optional["PreTrainedModel"] = None,
     **kwargs, # pylint: disable=W0613
 ):
-
     # 1. Set correct tokenizer and text encoder
     tokenizer: PreTrainedTokenizer = tokenizer or getattr(self, "tokenizer", None)
     text_encoder: PreTrainedModel  = text_encoder or getattr(self, "text_encoder", None)
     loaded_model_names_or_paths = {}
-
     assert tokenizer and text_encoder, 'Can not resolve `tokenizer` or `text_encoder`'
 
     # 2. Normalize inputs
@@ -99,9 +92,7 @@ def load_textual_inversion(
             if hasattr(component, "_hf_hook"):
                 is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)  # noqa: B009
                 is_sequential_cpu_offload = isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)  # noqa: B009
-                logger.info(
-                    "Accelerate hooks detected. Since you have called `load_textual_inversion()`, the previous hooks will be first removed. Then the textual inversion parameters will be loaded and the hooks will be applied again."
-                )
+                shared.log.debug("Accelerate hooks detected. Since you have called `load_textual_inversion()`, the previous hooks will be first removed. Then the textual inversion parameters will be loaded and the hooks will be applied again.")
                 remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
 
     # 7.2 save expected device and dtype

modules/textual_inversion/textual_inversion.py

@@ -1,3 +1,4 @@
+from typing import List, Optional, Union
 import csv
 import html
 import os
@@ -17,8 +18,8 @@ from modules.textual_inversion.ti_logging import save_settings_to_file
 from typing import List, Optional, Union
 from modules.files_cache import directory_files, directory_mtime, extension_filter
 
-TokenToAdd = namedtuple("TokenToAdd", ["clip_l", "clip_g"])
 
+TokenToAdd = namedtuple("TokenToAdd", ["clip_l", "clip_g"])
 TextualInversionTemplate = namedtuple("TextualInversionTemplate", ["name", "path"])
 textual_inversion_templates = {}
 

scripts/photomaker_model.py

@@ -221,12 +221,14 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
     ):
         device = device or self._execution_device
 
+        """
         if prompt is not None and isinstance(prompt, str):
             batch_size = 1
         elif prompt is not None and isinstance(prompt, list):
             batch_size = len(prompt)
         else:
             batch_size = prompt_embeds.shape[0]
+        """
 
         # Find the token id of the trigger word
         image_token_id = self.tokenizer_2.convert_tokens_to_ids(self.trigger_word)