Merge branch 'v2.3' into lstein/bugfix/improve-update-handling

Author: lstein

.github/CODEOWNERS

@@ -1,13 +1,13 @@
 # continuous integration
-/.github/workflows/ @mauwii @lstein @blessedcoolant
+/.github/workflows/  @lstein @blessedcoolant
 
 # documentation
-/docs/ @lstein @mauwii @blessedcoolant
-mkdocs.yml @mauwii @lstein
+/docs/ @lstein  @blessedcoolant
+mkdocs.yml  @lstein @ebr
 
 # installation and configuration
-/pyproject.toml @mauwii @lstein @ebr
-/docker/ @mauwii
+/pyproject.toml  @lstein @ebr
+/docker/ @lstein
 /scripts/ @ebr @lstein @blessedcoolant
 /installer/ @ebr @lstein 
 ldm/invoke/config @lstein @ebr
@@ -21,13 +21,13 @@ invokeai/configs @lstein @ebr  @blessedcoolant
 
 # generation and model management
 /ldm/*.py @lstein @blessedcoolant
-/ldm/generate.py @lstein @keturn
+/ldm/generate.py @lstein @gregghelt2
 /ldm/invoke/args.py @lstein @blessedcoolant
 /ldm/invoke/ckpt* @lstein @blessedcoolant
 /ldm/invoke/ckpt_generator @lstein @blessedcoolant
 /ldm/invoke/CLI.py @lstein @blessedcoolant
-/ldm/invoke/config @lstein @ebr @mauwii @blessedcoolant
-/ldm/invoke/generator @keturn @damian0815
+/ldm/invoke/config @lstein @ebr  @blessedcoolant
+/ldm/invoke/generator @gregghelt2 @damian0815
 /ldm/invoke/globals.py @lstein @blessedcoolant 
 /ldm/invoke/merge_diffusers.py @lstein @blessedcoolant
 /ldm/invoke/model_manager.py @lstein @blessedcoolant
@@ -36,17 +36,17 @@ invokeai/configs @lstein @ebr  @blessedcoolant
 /ldm/invoke/restoration @lstein @blessedcoolant
 
 # attention, textual inversion, model configuration
-/ldm/models @damian0815 @keturn @blessedcoolant
+/ldm/models @damian0815 @gregghelt2 @blessedcoolant
 /ldm/modules/textual_inversion_manager.py @lstein @blessedcoolant
-/ldm/modules/attention.py @damian0815 @keturn
-/ldm/modules/diffusionmodules @damian0815 @keturn
-/ldm/modules/distributions @damian0815 @keturn
-/ldm/modules/ema.py  @damian0815 @keturn
+/ldm/modules/attention.py @damian0815 @gregghelt2
+/ldm/modules/diffusionmodules @damian0815 @gregghelt2
+/ldm/modules/distributions @damian0815 @gregghelt2
+/ldm/modules/ema.py  @damian0815 @gregghelt2
 /ldm/modules/embedding_manager.py @lstein
-/ldm/modules/encoders @damian0815 @keturn
-/ldm/modules/image_degradation @damian0815 @keturn
-/ldm/modules/losses  @damian0815 @keturn
-/ldm/modules/x_transformer.py @damian0815 @keturn
+/ldm/modules/encoders @damian0815 @gregghelt2
+/ldm/modules/image_degradation @damian0815 @gregghelt2
+/ldm/modules/losses  @damian0815 @gregghelt2
+/ldm/modules/x_transformer.py @damian0815 @gregghelt2
 
 # Nodes
 apps/ @Kyle0654 @jpphoto

invokeai/backend/invoke_ai_web_server.py

@@ -30,7 +30,6 @@
     get_tokens_for_prompt_object,
     get_prompt_structure,
     split_weighted_subprompts,
-    get_tokenizer,
 )
 from ldm.invoke.generator.diffusers_pipeline import PipelineIntermediateState
 from ldm.invoke.generator.inpaint import infill_methods
@@ -1314,7 +1313,7 @@ def image_done(image, seed, first_seed, attention_maps_image=None):
                     None
                     if type(parsed_prompt) is Blend
                     else get_tokens_for_prompt_object(
-                        get_tokenizer(self.generate.model), parsed_prompt
+                        self.generate.model.tokenizer, parsed_prompt
                     )
                 )
                 attention_maps_image_base64_url = (

ldm/invoke/conditioning.py

@@ -15,19 +15,10 @@
 from compel.prompt_parser import FlattenedPrompt, Blend, Fragment, CrossAttentionControlSubstitute, PromptParser, \
     Conjunction
 from .devices import torch_dtype
+from .generator.diffusers_pipeline import StableDiffusionGeneratorPipeline
 from ..models.diffusion.shared_invokeai_diffusion import InvokeAIDiffuserComponent
 from ldm.invoke.globals import Globals
 
-def get_tokenizer(model) -> CLIPTokenizer:
-    # TODO remove legacy ckpt fallback handling
-    return (getattr(model, 'tokenizer', None) # diffusers
-            or model.cond_stage_model.tokenizer) # ldm
-
-def get_text_encoder(model) -> Any:
-    # TODO remove legacy ckpt fallback handling
-    return (getattr(model, 'text_encoder', None)  # diffusers
-            or UnsqueezingLDMTransformer(model.cond_stage_model.transformer)) # ldm
-
 class UnsqueezingLDMTransformer:
     def __init__(self, ldm_transformer):
         self.ldm_transformer = ldm_transformer
@@ -41,15 +32,15 @@ def __call__(self, *args, **kwargs):
         return insufficiently_unsqueezed_tensor.unsqueeze(0)
 
 
-def get_uc_and_c_and_ec(prompt_string, model, log_tokens=False, skip_normalize_legacy_blend=False):
+def get_uc_and_c_and_ec(prompt_string,
+                        model: StableDiffusionGeneratorPipeline,
+                        log_tokens=False, skip_normalize_legacy_blend=False):
     # lazy-load any deferred textual inversions.
     # this might take a couple of seconds the first time a textual inversion is used.
     model.textual_inversion_manager.create_deferred_token_ids_for_any_trigger_terms(prompt_string)
 
-    tokenizer = get_tokenizer(model)
-    text_encoder = get_text_encoder(model)
-    compel = Compel(tokenizer=tokenizer,
-                    text_encoder=text_encoder,
+    compel = Compel(tokenizer=model.tokenizer,
+                    text_encoder=model.text_encoder,
                     textual_inversion_manager=model.textual_inversion_manager,
                     dtype_for_device_getter=torch_dtype)
 
@@ -78,14 +69,20 @@ def get_uc_and_c_and_ec(prompt_string, model, log_tokens=False, skip_normalize_l
     negative_conjunction = Compel.parse_prompt_string(negative_prompt_string)
     negative_prompt: FlattenedPrompt | Blend = negative_conjunction.prompts[0]
 
+    tokens_count = get_max_token_count(model.tokenizer, positive_prompt)
     if log_tokens or getattr(Globals, "log_tokenization", False):
-        log_tokenization(positive_prompt, negative_prompt, tokenizer=tokenizer)
-
-    c, options = compel.build_conditioning_tensor_for_prompt_object(positive_prompt)
-    uc, _ = compel.build_conditioning_tensor_for_prompt_object(negative_prompt)
-
-    tokens_count = get_max_token_count(tokenizer, positive_prompt)
-
+        log_tokenization(positive_prompt, negative_prompt, tokenizer=model.tokenizer)
+
+    # some LoRA models also mess with the text encoder, so they must be active while compel builds conditioning tensors
+    lora_conditioning_ec = InvokeAIDiffuserComponent.ExtraConditioningInfo(tokens_count_including_eos_bos=tokens_count,
+                                                                                            lora_conditions=lora_conditions)
+    with InvokeAIDiffuserComponent.custom_attention_context(model.unet,
+                                                            extra_conditioning_info=lora_conditioning_ec,
+                                                            step_count=-1):
+        c, options = compel.build_conditioning_tensor_for_prompt_object(positive_prompt)
+        uc, _ = compel.build_conditioning_tensor_for_prompt_object(negative_prompt)
+
+    # now build the "real" ec
     ec = InvokeAIDiffuserComponent.ExtraConditioningInfo(tokens_count_including_eos_bos=tokens_count,
                                                          cross_attention_control_args=options.get(
                                                              'cross_attention_control', None),

ldm/invoke/config/model_install.py

@@ -196,16 +196,6 @@ def create(self):
             scroll_exit=True,
         )
         self.nextrely += 1
-        self.convert_models = self.add_widget_intelligent(
-            npyscreen.TitleSelectOne,
-            name="== CONVERT IMPORTED MODELS INTO DIFFUSERS==",
-            values=["Keep original format", "Convert to diffusers"],
-            value=0,
-            begin_entry_at=4,
-            max_height=4,
-            hidden=True,  # will appear when imported models box is edited
-            scroll_exit=True,
-        )
         self.cancel = self.add_widget_intelligent(
             npyscreen.ButtonPress,
             name="CANCEL",
@@ -240,8 +230,6 @@ def create(self):
             self.show_directory_fields.addVisibleWhenSelected(i)
 
         self.show_directory_fields.when_value_edited = self._clear_scan_directory
-        self.import_model_paths.when_value_edited = self._show_hide_convert
-        self.autoload_directory.when_value_edited = self._show_hide_convert
 
     def resize(self):
         super().resize()
@@ -252,13 +240,6 @@ def _clear_scan_directory(self):
         if not self.show_directory_fields.value:
             self.autoload_directory.value = ""
 
-    def _show_hide_convert(self):
-        model_paths = self.import_model_paths.value or ""
-        autoload_directory = self.autoload_directory.value or ""
-        self.convert_models.hidden = (
-            len(model_paths) == 0 and len(autoload_directory) == 0
-        )
-
     def _get_starter_model_labels(self) -> List[str]:
         window_width, window_height = get_terminal_size()
         label_width = 25
@@ -318,7 +299,6 @@ def marshall_arguments(self):
         .scan_directory: Path to a directory of models to scan and import
         .autoscan_on_startup:  True if invokeai should scan and import at startup time
         .import_model_paths:   list of URLs, repo_ids and file paths to import
-        .convert_to_diffusers: if True, convert legacy checkpoints into diffusers
         """
         # we're using a global here rather than storing the result in the parentapp
         # due to some bug in npyscreen that is causing attributes to be lost
@@ -354,7 +334,6 @@ def marshall_arguments(self):
 
         # URLs and the like
         selections.import_model_paths = self.import_model_paths.value.split()
-        selections.convert_to_diffusers = self.convert_models.value[0] == 1
 
 
 class AddModelApplication(npyscreen.NPSAppManaged):
@@ -367,7 +346,6 @@ def __init__(self):
             scan_directory=None,
             autoscan_on_startup=None,
             import_model_paths=None,
-            convert_to_diffusers=None,
         )
 
     def onStart(self):
@@ -387,15 +365,13 @@ def process_and_execute(opt: Namespace, selections: Namespace):
     directory_to_scan = selections.scan_directory
     scan_at_startup = selections.autoscan_on_startup
     potential_models_to_install = selections.import_model_paths
-    convert_to_diffusers = selections.convert_to_diffusers
 
     install_requested_models(
         install_initial_models=models_to_install,
         remove_models=models_to_remove,
         scan_directory=Path(directory_to_scan) if directory_to_scan else None,
         external_models=potential_models_to_install,
         scan_at_startup=scan_at_startup,
-        convert_to_diffusers=convert_to_diffusers,
         precision="float32"
         if opt.full_precision
         else choose_precision(torch.device(choose_torch_device())),

ldm/invoke/config/model_install_backend.py

@@ -68,7 +68,6 @@ def install_requested_models(
         scan_directory: Path = None,
         external_models: List[str] = None,
         scan_at_startup: bool = False,
-        convert_to_diffusers: bool = False,
         precision: str = "float16",
         purge_deleted: bool = False,
         config_file_path: Path = None,
@@ -111,20 +110,20 @@ def install_requested_models(
     if len(external_models)>0:
         print("== INSTALLING EXTERNAL MODELS ==")
         for path_url_or_repo in external_models:
+            print(f'DEBUG: path_url_or_repo = {path_url_or_repo}')
             try:
                 model_manager.heuristic_import(
                     path_url_or_repo,
-                    convert=convert_to_diffusers,
                     config_file_callback=_pick_configuration_file,
                     commit_to_conf=config_file_path
                 )
             except KeyboardInterrupt:
                 sys.exit(-1)
-            except Exception:
-                pass
+            except Exception as e:
+                print(f'An exception has occurred: {str(e)}')
 
     if scan_at_startup and scan_directory.is_dir():
-        argument = '--autoconvert' if convert_to_diffusers else '--autoimport'
+        argument = '--autoconvert'
         initfile = Path(Globals.root, Globals.initfile)
         replacement = Path(Globals.root, f'{Globals.initfile}.new')
         directory = str(scan_directory).replace('\\','/')

ldm/invoke/generator/diffusers_pipeline.py

@@ -467,8 +467,9 @@ def generate_latents_from_embeddings(self, latents: torch.Tensor, timesteps,
         if additional_guidance is None:
             additional_guidance = []
         extra_conditioning_info = conditioning_data.extra
-        with self.invokeai_diffuser.custom_attention_context(extra_conditioning_info=extra_conditioning_info,
-                                                             step_count=len(self.scheduler.timesteps)
+        with InvokeAIDiffuserComponent.custom_attention_context(self.invokeai_diffuser.model,
+                                                                extra_conditioning_info=extra_conditioning_info,
+                                                                step_count=len(self.scheduler.timesteps)
                                                              ):
 
             yield PipelineIntermediateState(run_id=run_id, step=-1, timestep=self.scheduler.num_train_timesteps,

ldm/models/diffusion/cross_attention_control.py

@@ -288,16 +288,7 @@ def get_invokeai_attention_mem_efficient(self, q, k, v):
         return self.einsum_op_tensor_mem(q, k, v, 32)
 
 
-
-def restore_default_cross_attention(model, is_running_diffusers: bool, processors_to_restore: Optional[AttnProcessor]=None):
-    if is_running_diffusers:
-        unet = model
-        unet.set_attn_processor(processors_to_restore or CrossAttnProcessor())
-    else:
-        remove_attention_function(model)
-
-
-def override_cross_attention(model, context: Context, is_running_diffusers = False):
+def setup_cross_attention_control_attention_processors(unet: UNet2DConditionModel, context: Context):
     """
     Inject attention parameters and functions into the passed in model to enable cross attention editing.
 
@@ -323,22 +314,15 @@ def override_cross_attention(model, context: Context, is_running_diffusers = Fal
 
     context.cross_attention_mask = mask.to(device)
     context.cross_attention_index_map = indices.to(device)
-    if is_running_diffusers:
-        unet = model
-        old_attn_processors = unet.attn_processors
-        if torch.backends.mps.is_available():
-            # see note in StableDiffusionGeneratorPipeline.__init__ about borked slicing on MPS
-            unet.set_attn_processor(SwapCrossAttnProcessor())
-        else:
-            # try to re-use an existing slice size
-            default_slice_size = 4
-            slice_size = next((p.slice_size for p in old_attn_processors.values() if type(p) is SlicedAttnProcessor), default_slice_size)
-            unet.set_attn_processor(SlicedSwapCrossAttnProcesser(slice_size=slice_size))
+    old_attn_processors = unet.attn_processors
+    if torch.backends.mps.is_available():
+        # see note in StableDiffusionGeneratorPipeline.__init__ about borked slicing on MPS
+        unet.set_attn_processor(SwapCrossAttnProcessor())
     else:
-        context.register_cross_attention_modules(model)
-        inject_attention_function(model, context)
-
-
+        # try to re-use an existing slice size
+        default_slice_size = 4
+        slice_size = next((p.slice_size for p in old_attn_processors.values() if type(p) is SlicedAttnProcessor), default_slice_size)
+        unet.set_attn_processor(SlicedSwapCrossAttnProcesser(slice_size=slice_size))
 
 
 def get_cross_attention_modules(model, which: CrossAttentionType) -> list[tuple[str, InvokeAICrossAttentionMixin]]:

ldm/models/diffusion/ddim.py

@@ -12,17 +12,6 @@ def __init__(self, model, schedule='linear', device=None, **kwargs):
         self.invokeai_diffuser = InvokeAIDiffuserComponent(self.model,
                                                            model_forward_callback = lambda x, sigma, cond: self.model.apply_model(x, sigma, cond))
 
-    def prepare_to_sample(self, t_enc, **kwargs):
-        super().prepare_to_sample(t_enc, **kwargs)
-
-        extra_conditioning_info = kwargs.get('extra_conditioning_info', None)
-        all_timesteps_count = kwargs.get('all_timesteps_count', t_enc)
-
-        if extra_conditioning_info is not None and extra_conditioning_info.wants_cross_attention_control:
-            self.invokeai_diffuser.override_attention_processors(extra_conditioning_info, step_count = all_timesteps_count)
-        else:
-            self.invokeai_diffuser.restore_default_cross_attention()
-
 
     # This is the central routine
     @torch.no_grad()

ldm/models/diffusion/ksampler.py

@@ -38,15 +38,6 @@ def __init__(self, model, threshold = 0, warmup = 0):
                                                            model_forward_callback=lambda x, sigma, cond: self.inner_model(x, sigma, cond=cond))
 
 
-    def prepare_to_sample(self, t_enc, **kwargs):
-
-        extra_conditioning_info = kwargs.get('extra_conditioning_info', None)
-
-        if extra_conditioning_info is not None and extra_conditioning_info.wants_cross_attention_control:
-            self.invokeai_diffuser.override_attention_processors(extra_conditioning_info, step_count = t_enc)
-        else:
-            self.invokeai_diffuser.restore_default_cross_attention()
-
 
     def forward(self, x, sigma, uncond, cond, cond_scale):
         next_x = self.invokeai_diffuser.do_diffusion_step(x, sigma, uncond, cond, cond_scale)

ldm/models/diffusion/plms.py

@@ -14,17 +14,6 @@ class PLMSSampler(Sampler):
     def __init__(self, model, schedule='linear', device=None, **kwargs):
         super().__init__(model,schedule,model.num_timesteps, device)
 
-    def prepare_to_sample(self, t_enc, **kwargs):
-        super().prepare_to_sample(t_enc, **kwargs)
-
-        extra_conditioning_info = kwargs.get('extra_conditioning_info', None)
-        all_timesteps_count = kwargs.get('all_timesteps_count', t_enc)
-
-        if extra_conditioning_info is not None and extra_conditioning_info.wants_cross_attention_control:
-            self.invokeai_diffuser.override_attention_processors(extra_conditioning_info, step_count = all_timesteps_count)
-        else:
-            self.invokeai_diffuser.restore_default_cross_attention()
-
 
     # this is the essential routine
     @torch.no_grad()