SDXL support

modules/launch_utils.py

@@ -224,6 +224,20 @@ def run_extensions_installers(settings_file):
         run_extension_installer(os.path.join(extensions_dir, dirname_extension))
 
 
+def mute_sdxl_imports():
+    """create fake modules that SDXL wants to import but doesn't actually use for our purposes"""
+
+    import importlib
+
+    module = importlib.util.module_from_spec(importlib.machinery.ModuleSpec('taming.modules.losses.lpips', None))
+    module.LPIPS = None
+    sys.modules['taming.modules.losses.lpips'] = module
+
+    module = importlib.util.module_from_spec(importlib.machinery.ModuleSpec('sgm.data', None))
+    module.StableDataModuleFromConfig = None
+    sys.modules['sgm.data'] = module
+
+
 def prepare_environment():
     torch_index_url = os.environ.get('TORCH_INDEX_URL', "https://download.pytorch.org/whl/cu118")
     torch_command = os.environ.get('TORCH_COMMAND', f"pip install torch==2.0.1 torchvision==0.15.2 --extra-index-url {torch_index_url}")
@@ -319,11 +333,14 @@ def prepare_environment():
     if args.update_all_extensions:
         git_pull_recursive(extensions_dir)
 
+    mute_sdxl_imports()
+
     if "--exit" in sys.argv:
         print("Exiting because of --exit argument")
         exit(0)
 
 
+
 def configure_for_tests():
     if "--api" not in sys.argv:
         sys.argv.append("--api")

modules/lowvram.py

@@ -53,19 +53,46 @@ def setup_for_low_vram(sd_model, use_medvram):
         send_me_to_gpu(first_stage_model, None)
         return first_stage_model_decode(z)
 
-    # for SD1, cond_stage_model is CLIP and its NN is in the tranformer frield, but for SD2, it's open clip, and it's in model field
+    to_remain_in_cpu = [
-    if hasattr(sd_model.cond_stage_model, 'model'):
+        (sd_model, 'first_stage_model'),
-        sd_model.cond_stage_model.transformer = sd_model.cond_stage_model.model
+        (sd_model, 'depth_model'),
+        (sd_model, 'embedder'),
+        (sd_model, 'model'),
+        (sd_model, 'embedder'),
+    ]
 
-    # remove several big modules: cond, first_stage, depth/embedder (if applicable), and unet from the model and then
+    is_sdxl = hasattr(sd_model, 'conditioner')
-    # send the model to GPU. Then put modules back. the modules will be in CPU.
+    is_sd2 = not is_sdxl and hasattr(sd_model.cond_stage_model, 'model')
-    stored = sd_model.cond_stage_model.transformer, sd_model.first_stage_model, getattr(sd_model, 'depth_model', None), getattr(sd_model, 'embedder', None), sd_model.model
+
-    sd_model.cond_stage_model.transformer, sd_model.first_stage_model, sd_model.depth_model, sd_model.embedder, sd_model.model = None, None, None, None, None
+    if is_sdxl:
+        to_remain_in_cpu.append((sd_model, 'conditioner'))
+    elif is_sd2:
+        to_remain_in_cpu.append((sd_model.cond_stage_model, 'model'))
+    else:
+        to_remain_in_cpu.append((sd_model.cond_stage_model, 'transformer'))
+
+    # remove several big modules: cond, first_stage, depth/embedder (if applicable), and unet from the model
+    stored = []
+    for obj, field in to_remain_in_cpu:
+        module = getattr(obj, field, None)
+        stored.append(module)
+        setattr(obj, field, None)
+
+    # send the model to GPU.
     sd_model.to(devices.device)
-    sd_model.cond_stage_model.transformer, sd_model.first_stage_model, sd_model.depth_model, sd_model.embedder, sd_model.model = stored
+
+    # put modules back. the modules will be in CPU.
+    for (obj, field), module in zip(to_remain_in_cpu, stored):
+        setattr(obj, field, module)
 
     # register hooks for those the first three models
+    if is_sdxl:
+        sd_model.conditioner.register_forward_pre_hook(send_me_to_gpu)
+    elif is_sd2:
+        sd_model.cond_stage_model.model.register_forward_pre_hook(send_me_to_gpu)
+    else:
         sd_model.cond_stage_model.transformer.register_forward_pre_hook(send_me_to_gpu)
+
     sd_model.first_stage_model.register_forward_pre_hook(send_me_to_gpu)
     sd_model.first_stage_model.encode = first_stage_model_encode_wrap
     sd_model.first_stage_model.decode = first_stage_model_decode_wrap
@@ -75,10 +102,6 @@ def setup_for_low_vram(sd_model, use_medvram):
         sd_model.embedder.register_forward_pre_hook(send_me_to_gpu)
     parents[sd_model.cond_stage_model.transformer] = sd_model.cond_stage_model
 
-    if hasattr(sd_model.cond_stage_model, 'model'):
-        sd_model.cond_stage_model.model = sd_model.cond_stage_model.transformer
-        del sd_model.cond_stage_model.transformer
-
     if use_medvram:
         sd_model.model.register_forward_pre_hook(send_me_to_gpu)
     else:

modules/paths.py

@@ -20,7 +20,7 @@ assert sd_path is not None, f"Couldn't find Stable Diffusion in any of: {possibl
 
 path_dirs = [
     (sd_path, 'ldm', 'Stable Diffusion', []),
-    (os.path.join(sd_path, '../generative-models'), 'sgm', 'Stable Diffusion XL', []),
+    (os.path.join(sd_path, '../generative-models'), 'sgm', 'Stable Diffusion XL', ["sgm"]),
     (os.path.join(sd_path, '../CodeFormer'), 'inference_codeformer.py', 'CodeFormer', []),
     (os.path.join(sd_path, '../BLIP'), 'models/blip.py', 'BLIP', []),
     (os.path.join(sd_path, '../k-diffusion'), 'k_diffusion/sampling.py', 'k_diffusion', ["atstart"]),
@@ -36,6 +36,13 @@ for d, must_exist, what, options in path_dirs:
         d = os.path.abspath(d)
         if "atstart" in options:
             sys.path.insert(0, d)
+        elif "sgm" in options:
+            # Stable Diffusion XL repo has scripts dir with __init__.py in it which ruins every extension's scripts dir, so we
+            # import sgm and remove it from sys.path so that when a script imports scripts.something, it doesbn't use sgm's scripts dir.
+
+            sys.path.insert(0, d)
+            import sgm
+            sys.path.pop(0)
         else:
             sys.path.append(d)
         paths[what] = d

modules/processing.py

@@ -343,10 +343,13 @@ class StableDiffusionProcessing:
         return cache[1]
 
     def setup_conds(self):
+        prompts = prompt_parser.SdConditioning(self.prompts, width=self.width, height=self.height)
+        negative_prompts = prompt_parser.SdConditioning(self.negative_prompts, width=self.width, height=self.height)
+
         sampler_config = sd_samplers.find_sampler_config(self.sampler_name)
         self.step_multiplier = 2 if sampler_config and sampler_config.options.get("second_order", False) else 1
-        self.uc = self.get_conds_with_caching(prompt_parser.get_learned_conditioning, self.negative_prompts, self.steps * self.step_multiplier, [self.cached_uc], self.extra_network_data)
+        self.uc = self.get_conds_with_caching(prompt_parser.get_learned_conditioning, negative_prompts, self.steps * self.step_multiplier, [self.cached_uc], self.extra_network_data)
-        self.c = self.get_conds_with_caching(prompt_parser.get_multicond_learned_conditioning, self.prompts, self.steps * self.step_multiplier, [self.cached_c], self.extra_network_data)
+        self.c = self.get_conds_with_caching(prompt_parser.get_multicond_learned_conditioning, prompts, self.steps * self.step_multiplier, [self.cached_c], self.extra_network_data)
 
     def parse_extra_network_prompts(self):
         self.prompts, self.extra_network_data = extra_networks.parse_prompts(self.prompts)

modules/prompt_parser.py

@@ -1,3 +1,5 @@
+from __future__ import annotations
+
 import re
 from collections import namedtuple
 from typing import List
@@ -109,7 +111,19 @@ def get_learned_conditioning_prompt_schedules(prompts, steps):
 ScheduledPromptConditioning = namedtuple("ScheduledPromptConditioning", ["end_at_step", "cond"])
 
 
-def get_learned_conditioning(model, prompts, steps):
+class SdConditioning(list):
+    """
+    A list with prompts for stable diffusion's conditioner model.
+    Can also specify width and height of created image - SDXL needs it.
+    """
+    def __init__(self, prompts, width=None, height=None):
+        super().__init__()
+        self.extend(prompts)
+        self.width = width or getattr(prompts, 'width', None)
+        self.height = height or getattr(prompts, 'height', None)
+
+
+def get_learned_conditioning(model, prompts: SdConditioning | list[str], steps):
     """converts a list of prompts into a list of prompt schedules - each schedule is a list of ScheduledPromptConditioning, specifying the comdition (cond),
     and the sampling step at which this condition is to be replaced by the next one.
 
@@ -160,11 +174,13 @@ def get_learned_conditioning(model, prompts, steps):
 re_AND = re.compile(r"\bAND\b")
 re_weight = re.compile(r"^(.*?)(?:\s*:\s*([-+]?(?:\d+\.?|\d*\.\d+)))?\s*$")
 
-def get_multicond_prompt_list(prompts):
+
+def get_multicond_prompt_list(prompts: SdConditioning | list[str]):
     res_indexes = []
 
-    prompt_flat_list = []
     prompt_indexes = {}
+    prompt_flat_list = SdConditioning(prompts)
+    prompt_flat_list.clear()
 
     for prompt in prompts:
         subprompts = re_AND.split(prompt)
@@ -201,6 +217,7 @@ class MulticondLearnedConditioning:
         self.shape: tuple = shape  # the shape field is needed to send this object to DDIM/PLMS
         self.batch: List[List[ComposableScheduledPromptConditioning]] = batch
 
+
 def get_multicond_learned_conditioning(model, prompts, steps) -> MulticondLearnedConditioning:
     """same as get_learned_conditioning, but returns a list of ScheduledPromptConditioning along with the weight objects for each prompt.
     For each prompt, the list is obtained by splitting the prompt using the AND separator.

modules/sd_hijack.py

@@ -15,6 +15,11 @@ import ldm.models.diffusion.ddim
 import ldm.models.diffusion.plms
 import ldm.modules.encoders.modules
 
+import sgm.modules.attention
+import sgm.modules.diffusionmodules.model
+import sgm.modules.diffusionmodules.openaimodel
+import sgm.modules.encoders.modules
+
 attention_CrossAttention_forward = ldm.modules.attention.CrossAttention.forward
 diffusionmodules_model_nonlinearity = ldm.modules.diffusionmodules.model.nonlinearity
 diffusionmodules_model_AttnBlock_forward = ldm.modules.diffusionmodules.model.AttnBlock.forward
@@ -56,6 +61,9 @@ def apply_optimizations(option=None):
     ldm.modules.diffusionmodules.model.nonlinearity = silu
     ldm.modules.diffusionmodules.openaimodel.th = sd_hijack_unet.th
 
+    sgm.modules.diffusionmodules.model.nonlinearity = silu
+    sgm.modules.diffusionmodules.openaimodel.th = sd_hijack_unet.th
+
     if current_optimizer is not None:
         current_optimizer.undo()
         current_optimizer = None
@@ -89,6 +97,10 @@ def undo_optimizations():
     ldm.modules.attention.CrossAttention.forward = hypernetwork.attention_CrossAttention_forward
     ldm.modules.diffusionmodules.model.AttnBlock.forward = diffusionmodules_model_AttnBlock_forward
 
+    sgm.modules.diffusionmodules.model.nonlinearity = diffusionmodules_model_nonlinearity
+    sgm.modules.attention.CrossAttention.forward = hypernetwork.attention_CrossAttention_forward
+    sgm.modules.diffusionmodules.model.AttnBlock.forward = diffusionmodules_model_AttnBlock_forward
+
 
 def fix_checkpoint():
     """checkpoints are now added and removed in embedding/hypernet code, since torch doesn't want
@@ -170,10 +182,19 @@ class StableDiffusionModelHijack:
         if conditioner:
             for i in range(len(conditioner.embedders)):
                 embedder = conditioner.embedders[i]
-                if type(embedder).__name__ == 'FrozenOpenCLIPEmbedder':
+                typename = type(embedder).__name__
+                if typename == 'FrozenOpenCLIPEmbedder':
                     embedder.model.token_embedding = EmbeddingsWithFixes(embedder.model.token_embedding, self)
                     m.cond_stage_model = sd_hijack_open_clip.FrozenOpenCLIPEmbedderWithCustomWords(embedder, self)
                     conditioner.embedders[i] = m.cond_stage_model
+                if typename == 'FrozenCLIPEmbedder':
+                    model_embeddings = m.cond_stage_model.transformer.text_model.embeddings
+                    model_embeddings.token_embedding = EmbeddingsWithFixes(model_embeddings.token_embedding, self)
+                    m.cond_stage_model = sd_hijack_clip.FrozenCLIPEmbedderWithCustomWords(embedder, self)
+                    conditioner.embedders[i] = m.cond_stage_model
+                if typename == 'FrozenOpenCLIPEmbedder2':
+                    embedder.model.token_embedding = EmbeddingsWithFixes(embedder.model.token_embedding, self)
+                    conditioner.embedders[i] = sd_hijack_open_clip.FrozenOpenCLIPEmbedder2WithCustomWords(embedder, self)
 
         if type(m.cond_stage_model) == xlmr.BertSeriesModelWithTransformation:
             model_embeddings = m.cond_stage_model.roberta.embeddings

modules/sd_hijack_clip.py

@@ -42,6 +42,10 @@ class FrozenCLIPEmbedderWithCustomWordsBase(torch.nn.Module):
         self.hijack: sd_hijack.StableDiffusionModelHijack = hijack
         self.chunk_length = 75
 
+        self.is_trainable = getattr(wrapped, 'is_trainable', False)
+        self.input_key = getattr(wrapped, 'input_key', 'txt')
+        self.legacy_ucg_val = None
+
     def empty_chunk(self):
         """creates an empty PromptChunk and returns it"""
 
@@ -199,8 +203,9 @@ class FrozenCLIPEmbedderWithCustomWordsBase(torch.nn.Module):
         """
         Accepts an array of texts; Passes texts through transformers network to create a tensor with numerical representation of those texts.
         Returns a tensor with shape of (B, T, C), where B is length of the array; T is length, in tokens, of texts (including padding) - T will
-        be a multiple of 77; and C is dimensionality of each token - for SD1 it's 768, and for SD2 it's 1024.
+        be a multiple of 77; and C is dimensionality of each token - for SD1 it's 768, for SD2 it's 1024, and for SDXL it's 1280.
         An example shape returned by this function can be: (2, 77, 768).
+        For SDXL, instead of returning one tensor avobe, it returns a tuple with two: the other one with shape (B, 1280) with pooled values.
         Webui usually sends just one text at a time through this function - the only time when texts is an array with more than one elemenet
         is when you do prompt editing: "a picture of a [cat:dog:0.4] eating ice cream"
         """
@@ -233,6 +238,9 @@ class FrozenCLIPEmbedderWithCustomWordsBase(torch.nn.Module):
             embeddings_list = ", ".join([f'{name} [{embedding.checksum()}]' for name, embedding in used_embeddings.items()])
             self.hijack.comments.append(f"Used embeddings: {embeddings_list}")
 
+        if getattr(self.wrapped, 'return_pooled', False):
+            return torch.hstack(zs), zs[0].pooled
+        else:
             return torch.hstack(zs)
 
     def process_tokens(self, remade_batch_tokens, batch_multipliers):
@@ -256,9 +264,9 @@ class FrozenCLIPEmbedderWithCustomWordsBase(torch.nn.Module):
         # restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise
         batch_multipliers = torch.asarray(batch_multipliers).to(devices.device)
         original_mean = z.mean()
-        z = z * batch_multipliers.reshape(batch_multipliers.shape + (1,)).expand(z.shape)
+        z *= batch_multipliers.reshape(batch_multipliers.shape + (1,)).expand(z.shape)
         new_mean = z.mean()
-        z = z * (original_mean / new_mean)
+        z *= (original_mean / new_mean)
 
         return z
 

modules/sd_hijack_open_clip.py

@@ -16,10 +16,6 @@ class FrozenOpenCLIPEmbedderWithCustomWords(sd_hijack_clip.FrozenCLIPEmbedderWit
         self.id_end = tokenizer.encoder["<end_of_text>"]
         self.id_pad = 0
 
-        self.is_trainable = getattr(wrapped, 'is_trainable', False)
-        self.input_key = getattr(wrapped, 'input_key', 'txt')
-        self.legacy_ucg_val = None
-
     def tokenize(self, texts):
         assert not opts.use_old_emphasis_implementation, 'Old emphasis implementation not supported for Open Clip'
 
@@ -39,3 +35,37 @@ class FrozenOpenCLIPEmbedderWithCustomWords(sd_hijack_clip.FrozenCLIPEmbedderWit
         embedded = self.wrapped.model.token_embedding.wrapped(ids).squeeze(0)
 
         return embedded
+
+
+class FrozenOpenCLIPEmbedder2WithCustomWords(sd_hijack_clip.FrozenCLIPEmbedderWithCustomWordsBase):
+    def __init__(self, wrapped, hijack):
+        super().__init__(wrapped, hijack)
+
+        self.comma_token = [v for k, v in tokenizer.encoder.items() if k == ',</w>'][0]
+        self.id_start = tokenizer.encoder["<start_of_text>"]
+        self.id_end = tokenizer.encoder["<end_of_text>"]
+        self.id_pad = 0
+
+    def tokenize(self, texts):
+        assert not opts.use_old_emphasis_implementation, 'Old emphasis implementation not supported for Open Clip'
+
+        tokenized = [tokenizer.encode(text) for text in texts]
+
+        return tokenized
+
+    def encode_with_transformers(self, tokens):
+        d = self.wrapped.encode_with_transformer(tokens)
+        z = d[self.wrapped.layer]
+
+        pooled = d.get("pooled")
+        if pooled is not None:
+            z.pooled = pooled
+
+        return z
+
+    def encode_embedding_init_text(self, init_text, nvpt):
+        ids = tokenizer.encode(init_text)
+        ids = torch.asarray([ids], device=devices.device, dtype=torch.int)
+        embedded = self.wrapped.model.token_embedding.wrapped(ids).squeeze(0)
+
+        return embedded

modules/sd_hijack_optimizations.py

@@ -14,7 +14,11 @@ from modules.hypernetworks import hypernetwork
 import ldm.modules.attention
 import ldm.modules.diffusionmodules.model
 
+import sgm.modules.attention
+import sgm.modules.diffusionmodules.model
+
 diffusionmodules_model_AttnBlock_forward = ldm.modules.diffusionmodules.model.AttnBlock.forward
+sgm_diffusionmodules_model_AttnBlock_forward = sgm.modules.diffusionmodules.model.AttnBlock.forward
 
 
 class SdOptimization:
@@ -39,6 +43,9 @@ class SdOptimization:
         ldm.modules.attention.CrossAttention.forward = hypernetwork.attention_CrossAttention_forward
         ldm.modules.diffusionmodules.model.AttnBlock.forward = diffusionmodules_model_AttnBlock_forward
 
+        sgm.modules.attention.CrossAttention.forward = hypernetwork.attention_CrossAttention_forward
+        sgm.modules.diffusionmodules.model.AttnBlock.forward = sgm_diffusionmodules_model_AttnBlock_forward
+
 
 class SdOptimizationXformers(SdOptimization):
     name = "xformers"
@@ -51,6 +58,8 @@ class SdOptimizationXformers(SdOptimization):
     def apply(self):
         ldm.modules.attention.CrossAttention.forward = xformers_attention_forward
         ldm.modules.diffusionmodules.model.AttnBlock.forward = xformers_attnblock_forward
+        sgm.modules.attention.CrossAttention.forward = xformers_attention_forward
+        sgm.modules.diffusionmodules.model.AttnBlock.forward = xformers_attnblock_forward
 
 
 class SdOptimizationSdpNoMem(SdOptimization):
@@ -65,6 +74,8 @@ class SdOptimizationSdpNoMem(SdOptimization):
     def apply(self):
         ldm.modules.attention.CrossAttention.forward = scaled_dot_product_no_mem_attention_forward
         ldm.modules.diffusionmodules.model.AttnBlock.forward = sdp_no_mem_attnblock_forward
+        sgm.modules.attention.CrossAttention.forward = scaled_dot_product_no_mem_attention_forward
+        sgm.modules.diffusionmodules.model.AttnBlock.forward = sdp_no_mem_attnblock_forward
 
 
 class SdOptimizationSdp(SdOptimizationSdpNoMem):
@@ -76,6 +87,8 @@ class SdOptimizationSdp(SdOptimizationSdpNoMem):
     def apply(self):
         ldm.modules.attention.CrossAttention.forward = scaled_dot_product_attention_forward
         ldm.modules.diffusionmodules.model.AttnBlock.forward = sdp_attnblock_forward
+        sgm.modules.attention.CrossAttention.forward = scaled_dot_product_attention_forward
+        sgm.modules.diffusionmodules.model.AttnBlock.forward = sdp_attnblock_forward
 
 
 class SdOptimizationSubQuad(SdOptimization):
@@ -86,6 +99,8 @@ class SdOptimizationSubQuad(SdOptimization):
     def apply(self):
         ldm.modules.attention.CrossAttention.forward = sub_quad_attention_forward
         ldm.modules.diffusionmodules.model.AttnBlock.forward = sub_quad_attnblock_forward
+        sgm.modules.attention.CrossAttention.forward = sub_quad_attention_forward
+        sgm.modules.diffusionmodules.model.AttnBlock.forward = sub_quad_attnblock_forward
 
 
 class SdOptimizationV1(SdOptimization):
@@ -94,9 +109,9 @@ class SdOptimizationV1(SdOptimization):
     cmd_opt = "opt_split_attention_v1"
     priority = 10
 
-
     def apply(self):
         ldm.modules.attention.CrossAttention.forward = split_cross_attention_forward_v1
+        sgm.modules.attention.CrossAttention.forward = split_cross_attention_forward_v1
 
 
 class SdOptimizationInvokeAI(SdOptimization):
@@ -109,6 +124,7 @@ class SdOptimizationInvokeAI(SdOptimization):
 
     def apply(self):
         ldm.modules.attention.CrossAttention.forward = split_cross_attention_forward_invokeAI
+        sgm.modules.attention.CrossAttention.forward = split_cross_attention_forward_invokeAI
 
 
 class SdOptimizationDoggettx(SdOptimization):
@@ -119,6 +135,8 @@ class SdOptimizationDoggettx(SdOptimization):
     def apply(self):
         ldm.modules.attention.CrossAttention.forward = split_cross_attention_forward
         ldm.modules.diffusionmodules.model.AttnBlock.forward = cross_attention_attnblock_forward
+        sgm.modules.attention.CrossAttention.forward = split_cross_attention_forward
+        sgm.modules.diffusionmodules.model.AttnBlock.forward = cross_attention_attnblock_forward
 
 
 def list_optimizers(res):
@@ -155,7 +173,7 @@ def get_available_vram():
 
 
 # see https://github.com/basujindal/stable-diffusion/pull/117 for discussion
-def split_cross_attention_forward_v1(self, x, context=None, mask=None):
+def split_cross_attention_forward_v1(self, x, context=None, mask=None, additional_tokens=None, n_times_crossframe_attn_in_self=0):
     h = self.heads
 
     q_in = self.to_q(x)
@@ -196,7 +214,7 @@ def split_cross_attention_forward_v1(self, x, context=None, mask=None):
 
 
 # taken from https://github.com/Doggettx/stable-diffusion and modified
-def split_cross_attention_forward(self, x, context=None, mask=None):
+def split_cross_attention_forward(self, x, context=None, mask=None, additional_tokens=None, n_times_crossframe_attn_in_self=0):
     h = self.heads
 
     q_in = self.to_q(x)
@@ -262,11 +280,13 @@ def split_cross_attention_forward(self, x, context=None, mask=None):
 # -- Taken from https://github.com/invoke-ai/InvokeAI and modified --
 mem_total_gb = psutil.virtual_memory().total // (1 << 30)
 
+
 def einsum_op_compvis(q, k, v):
     s = einsum('b i d, b j d -> b i j', q, k)
     s = s.softmax(dim=-1, dtype=s.dtype)
     return einsum('b i j, b j d -> b i d', s, v)
 
+
 def einsum_op_slice_0(q, k, v, slice_size):
     r = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype)
     for i in range(0, q.shape[0], slice_size):
@@ -274,6 +294,7 @@ def einsum_op_slice_0(q, k, v, slice_size):
         r[i:end] = einsum_op_compvis(q[i:end], k[i:end], v[i:end])
     return r
 
+
 def einsum_op_slice_1(q, k, v, slice_size):
     r = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype)
     for i in range(0, q.shape[1], slice_size):
@@ -281,6 +302,7 @@ def einsum_op_slice_1(q, k, v, slice_size):
         r[:, i:end] = einsum_op_compvis(q[:, i:end], k, v)
     return r
 
+
 def einsum_op_mps_v1(q, k, v):
     if q.shape[0] * q.shape[1] <= 2**16: # (512x512) max q.shape[1]: 4096
         return einsum_op_compvis(q, k, v)
@@ -290,12 +312,14 @@ def einsum_op_mps_v1(q, k, v):
             slice_size -= 1
         return einsum_op_slice_1(q, k, v, slice_size)
 
+
 def einsum_op_mps_v2(q, k, v):
     if mem_total_gb > 8 and q.shape[0] * q.shape[1] <= 2**16:
         return einsum_op_compvis(q, k, v)
     else:
         return einsum_op_slice_0(q, k, v, 1)
 
+
 def einsum_op_tensor_mem(q, k, v, max_tensor_mb):
     size_mb = q.shape[0] * q.shape[1] * k.shape[1] * q.element_size() // (1 << 20)
     if size_mb <= max_tensor_mb:
@@ -305,6 +329,7 @@ def einsum_op_tensor_mem(q, k, v, max_tensor_mb):
         return einsum_op_slice_0(q, k, v, q.shape[0] // div)
     return einsum_op_slice_1(q, k, v, max(q.shape[1] // div, 1))
 
+
 def einsum_op_cuda(q, k, v):
     stats = torch.cuda.memory_stats(q.device)
     mem_active = stats['active_bytes.all.current']
@@ -315,6 +340,7 @@ def einsum_op_cuda(q, k, v):
     # Divide factor of safety as there's copying and fragmentation
     return einsum_op_tensor_mem(q, k, v, mem_free_total / 3.3 / (1 << 20))
 
+
 def einsum_op(q, k, v):
     if q.device.type == 'cuda':
         return einsum_op_cuda(q, k, v)
@@ -328,7 +354,8 @@ def einsum_op(q, k, v):
     # Tested on i7 with 8MB L3 cache.
     return einsum_op_tensor_mem(q, k, v, 32)
 
-def split_cross_attention_forward_invokeAI(self, x, context=None, mask=None):
+
+def split_cross_attention_forward_invokeAI(self, x, context=None, mask=None, additional_tokens=None, n_times_crossframe_attn_in_self=0):
     h = self.heads
 
     q = self.to_q(x)
@@ -356,7 +383,7 @@ def split_cross_attention_forward_invokeAI(self, x, context=None, mask=None):
 
 # Based on Birch-san's modified implementation of sub-quadratic attention from https://github.com/Birch-san/diffusers/pull/1
 # The sub_quad_attention_forward function is under the MIT License listed under Memory Efficient Attention in the Licenses section of the web UI interface
-def sub_quad_attention_forward(self, x, context=None, mask=None):
+def sub_quad_attention_forward(self, x, context=None, mask=None, additional_tokens=None, n_times_crossframe_attn_in_self=0):
     assert mask is None, "attention-mask not currently implemented for SubQuadraticCrossAttnProcessor."
 
     h = self.heads
@@ -392,6 +419,7 @@ def sub_quad_attention_forward(self, x, context=None, mask=None):
 
     return x
 
+
 def sub_quad_attention(q, k, v, q_chunk_size=1024, kv_chunk_size=None, kv_chunk_size_min=None, chunk_threshold=None, use_checkpoint=True):
     bytes_per_token = torch.finfo(q.dtype).bits//8
     batch_x_heads, q_tokens, _ = q.shape
@@ -442,7 +470,7 @@ def get_xformers_flash_attention_op(q, k, v):
     return None
 
 
-def xformers_attention_forward(self, x, context=None, mask=None):
+def xformers_attention_forward(self, x, context=None, mask=None, additional_tokens=None, n_times_crossframe_attn_in_self=0):
     h = self.heads
     q_in = self.to_q(x)
     context = default(context, x)
@@ -465,9 +493,10 @@ def xformers_attention_forward(self, x, context=None, mask=None):
     out = rearrange(out, 'b n h d -> b n (h d)', h=h)
     return self.to_out(out)
 
+
 # Based on Diffusers usage of scaled dot product attention from https://github.com/huggingface/diffusers/blob/c7da8fd23359a22d0df2741688b5b4f33c26df21/src/diffusers/models/cross_attention.py
 # The scaled_dot_product_attention_forward function contains parts of code under Apache-2.0 license listed under Scaled Dot Product Attention in the Licenses section of the web UI interface
-def scaled_dot_product_attention_forward(self, x, context=None, mask=None):
+def scaled_dot_product_attention_forward(self, x, context=None, mask=None, additional_tokens=None, n_times_crossframe_attn_in_self=0):
     batch_size, sequence_length, inner_dim = x.shape
 
     if mask is not None:
@@ -507,10 +536,12 @@ def scaled_dot_product_attention_forward(self, x, context=None, mask=None):
     hidden_states = self.to_out[1](hidden_states)
     return hidden_states
 
-def scaled_dot_product_no_mem_attention_forward(self, x, context=None, mask=None):
+
+def scaled_dot_product_no_mem_attention_forward(self, x, context=None, mask=None, additional_tokens=None, n_times_crossframe_attn_in_self=0):
     with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=True, enable_mem_efficient=False):
         return scaled_dot_product_attention_forward(self, x, context, mask)
 
+
 def cross_attention_attnblock_forward(self, x):
         h_ = x
         h_ = self.norm(h_)
@@ -569,6 +600,7 @@ def cross_attention_attnblock_forward(self, x):
 
         return h3
 
+
 def xformers_attnblock_forward(self, x):
     try:
         h_ = x
@@ -592,6 +624,7 @@ def xformers_attnblock_forward(self, x):
     except NotImplementedError:
         return cross_attention_attnblock_forward(self, x)
 
+
 def sdp_attnblock_forward(self, x):
     h_ = x
     h_ = self.norm(h_)
@@ -612,10 +645,12 @@ def sdp_attnblock_forward(self, x):
     out = self.proj_out(out)
     return x + out
 
+
 def sdp_no_mem_attnblock_forward(self, x):
     with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=True, enable_mem_efficient=False):
         return sdp_attnblock_forward(self, x)
 
+
 def sub_quad_attnblock_forward(self, x):
     h_ = x
     h_ = self.norm(h_)

modules/sd_models.py

@@ -411,6 +411,7 @@ def repair_config(sd_config):
 
 sd1_clip_weight = 'cond_stage_model.transformer.text_model.embeddings.token_embedding.weight'
 sd2_clip_weight = 'cond_stage_model.model.transformer.resblocks.0.attn.in_proj_weight'
+sdxl_clip_weight = 'conditioner.embedders.1.model.ln_final.weight'
 
 
 class SdModelData:
@@ -445,6 +446,15 @@ class SdModelData:
 model_data = SdModelData()
 
 
+def get_empty_cond(sd_model):
+    if hasattr(sd_model, 'conditioner'):
+        d = sd_model.get_learned_conditioning([""])
+        return d['crossattn']
+    else:
+        return sd_model.cond_stage_model([""])
+
+
+
 def load_model(checkpoint_info=None, already_loaded_state_dict=None):
     from modules import lowvram, sd_hijack
     checkpoint_info = checkpoint_info or select_checkpoint()
@@ -465,7 +475,7 @@ def load_model(checkpoint_info=None, already_loaded_state_dict=None):
         state_dict = get_checkpoint_state_dict(checkpoint_info, timer)
 
     checkpoint_config = sd_models_config.find_checkpoint_config(state_dict, checkpoint_info)
-    clip_is_included_into_sd = sd1_clip_weight in state_dict or sd2_clip_weight in state_dict
+    clip_is_included_into_sd = sd1_clip_weight in state_dict or sd2_clip_weight in state_dict or sdxl_clip_weight in state_dict
 
     timer.record("find config")
 
@@ -517,7 +527,7 @@ def load_model(checkpoint_info=None, already_loaded_state_dict=None):
     timer.record("scripts callbacks")
 
     with devices.autocast(), torch.no_grad():
-        sd_model.cond_stage_model_empty_prompt = sd_model.cond_stage_model([""])
+        sd_model.cond_stage_model_empty_prompt = get_empty_cond(sd_model)
 
     timer.record("calculate empty prompt")
 

modules/sd_models_config.py

@@ -14,6 +14,7 @@ config_sd2 = os.path.join(sd_repo_configs_path, "v2-inference.yaml")
 config_sd2v = os.path.join(sd_repo_configs_path, "v2-inference-v.yaml")
 config_sd2v = os.path.join(sd_xl_repo_configs_path, "sd_2_1_768.yaml")
 config_sd2_inpainting = os.path.join(sd_repo_configs_path, "v2-inpainting-inference.yaml")
+config_sdxl = os.path.join(sd_xl_repo_configs_path, "sd_xl_base.yaml")
 config_depth_model = os.path.join(sd_repo_configs_path, "v2-midas-inference.yaml")
 config_unclip = os.path.join(sd_repo_configs_path, "v2-1-stable-unclip-l-inference.yaml")
 config_unopenclip = os.path.join(sd_repo_configs_path, "v2-1-stable-unclip-h-inference.yaml")
@@ -70,7 +71,9 @@ def guess_model_config_from_state_dict(sd, filename):
     diffusion_model_input = sd.get('model.diffusion_model.input_blocks.0.0.weight', None)
     sd2_variations_weight = sd.get('embedder.model.ln_final.weight', None)
 
-    if sd.get('depth_model.model.pretrained.act_postprocess3.0.project.0.bias', None) is not None:
+    if sd.get('conditioner.embedders.1.model.ln_final.weight', None) is not None:
+        return config_sdxl
+    elif sd.get('depth_model.model.pretrained.act_postprocess3.0.project.0.bias', None) is not None:
         return config_depth_model
     elif sd2_variations_weight is not None and sd2_variations_weight.shape[0] == 768:
         return config_unclip

modules/sd_models_xl.py

@@ -1,18 +1,30 @@
 from __future__ import annotations
 
+import sys
+
 import torch
 
 import sgm.models.diffusion
 import sgm.modules.diffusionmodules.denoiser_scaling
 import sgm.modules.diffusionmodules.discretizer
-from modules import devices
+from modules import devices, shared, prompt_parser
 
 
-def get_learned_conditioning(self: sgm.models.diffusion.DiffusionEngine, batch: list[str]):
+def get_learned_conditioning(self: sgm.models.diffusion.DiffusionEngine, batch: prompt_parser.SdConditioning | list[str]):
     for embedder in self.conditioner.embedders:
         embedder.ucg_rate = 0.0
 
-    c = self.conditioner({'txt': batch})
+    width = getattr(self, 'target_width', 1024)
+    height = getattr(self, 'target_height', 1024)
+
+    sdxl_conds = {
+        "txt": batch,
+        "original_size_as_tuple": torch.tensor([height, width]).repeat(len(batch), 1).to(devices.device, devices.dtype),
+        "crop_coords_top_left": torch.tensor([shared.opts.sdxl_crop_top, shared.opts.sdxl_crop_left]).repeat(len(batch), 1).to(devices.device, devices.dtype),
+        "target_size_as_tuple": torch.tensor([height, width]).repeat(len(batch), 1).to(devices.device, devices.dtype),
+    }
+
+    c = self.conditioner(sdxl_conds)
 
     return c
 
@@ -26,7 +38,7 @@ def extend_sdxl(model):
     model.model.diffusion_model.dtype = dtype
     model.model.conditioning_key = 'crossattn'
 
-    model.cond_stage_model = [x for x in model.conditioner.embedders if type(x).__name__ == 'FrozenOpenCLIPEmbedder'][0]
+    model.cond_stage_model = [x for x in model.conditioner.embedders if 'CLIPEmbedder' in type(x).__name__][0]
     model.cond_stage_key = model.cond_stage_model.input_key
 
     model.parameterization = "v" if isinstance(model.denoiser.scaling, sgm.modules.diffusionmodules.denoiser_scaling.VScaling) else "eps"
@@ -34,7 +46,14 @@ def extend_sdxl(model):
     discretization = sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization()
     model.alphas_cumprod = torch.asarray(discretization.alphas_cumprod, device=devices.device, dtype=dtype)
 
+    model.is_xl = True
+
 
 sgm.models.diffusion.DiffusionEngine.get_learned_conditioning = get_learned_conditioning
 sgm.models.diffusion.DiffusionEngine.apply_model = apply_model
 
+sgm.modules.attention.print = lambda *args: None
+sgm.modules.diffusionmodules.model.print = lambda *args: None
+sgm.modules.diffusionmodules.openaimodel.print = lambda *args: None
+sgm.modules.encoders.modules.print = lambda *args: None
+

modules/sd_samplers_kdiffusion.py

@@ -186,7 +186,7 @@ class CFGDenoiser(torch.nn.Module):
                 for batch_offset in range(0, x_out.shape[0], batch_size):
                     a = batch_offset
                     b = a + batch_size
-                    x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond=make_condition_dict(cond_in[a:b], image_cond_in[a:b]))
+                    x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond=make_condition_dict(subscript_cond(cond_in, a, b), image_cond_in[a:b]))
         else:
             x_out = torch.zeros_like(x_in)
             batch_size = batch_size*2 if shared.batch_cond_uncond else batch_size

modules/shared.py

@@ -428,6 +428,8 @@ options_templates.update(options_section(('sd', "Stable Diffusion"), {
     "CLIP_stop_at_last_layers": OptionInfo(1, "Clip skip", gr.Slider, {"minimum": 1, "maximum": 12, "step": 1}).link("wiki", "https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#clip-skip").info("ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer"),
     "upcast_attn": OptionInfo(False, "Upcast cross attention layer to float32"),
     "randn_source": OptionInfo("GPU", "Random number generator source.", gr.Radio, {"choices": ["GPU", "CPU"]}).info("changes seeds drastically; use CPU to produce the same picture across different videocard vendors"),
+    "sdxl_crop_top": OptionInfo(0, "SDXL top coordinate of the crop"),
+    "sdxl_crop_left": OptionInfo(0, "SDXL left coordinate of the crop"),
 }))
 
 options_templates.update(options_section(('optimizations', "Optimizations"), {

requirements.txt

@@ -14,6 +14,7 @@ kornia
 lark
 numpy
 omegaconf
+open-clip-torch
 
 piexif
 psutil

requirements_versions.txt

@@ -15,6 +15,7 @@ kornia==0.6.7
 lark==1.1.2
 numpy==1.23.5
 omegaconf==2.2.3
+open-clip-torch==2.20.0
 piexif==1.1.3
 psutil~=5.9.5
 pytorch_lightning==1.9.4