diff options
Diffstat (limited to 'modules/sd_hijack.py')
| -rw-r--r-- | modules/sd_hijack.py | 288 |
1 files changed, 81 insertions, 207 deletions
diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py index d68f89cc..b824b5bf 100644 --- a/modules/sd_hijack.py +++ b/modules/sd_hijack.py @@ -8,25 +8,54 @@ from torch import einsum from torch.nn.functional import silu
import modules.textual_inversion.textual_inversion
-from modules import prompt_parser, devices, sd_hijack_optimizations, shared, hypernetwork
+from modules import prompt_parser, devices, sd_hijack_optimizations, shared, sd_hijack_checkpoint
+from modules.hypernetworks import hypernetwork
from modules.shared import opts, device, cmd_opts
+from modules import sd_hijack_clip, sd_hijack_open_clip
+
+from modules.sd_hijack_optimizations import invokeAI_mps_available
import ldm.modules.attention
import ldm.modules.diffusionmodules.model
+import ldm.models.diffusion.ddim
+import ldm.models.diffusion.plms
+import ldm.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
+# new memory efficient cross attention blocks do not support hypernets and we already
+# have memory efficient cross attention anyway, so this disables SD2.0's memory efficient cross attention
+ldm.modules.attention.MemoryEfficientCrossAttention = ldm.modules.attention.CrossAttention
+ldm.modules.attention.BasicTransformerBlock.ATTENTION_MODES["softmax-xformers"] = ldm.modules.attention.CrossAttention
+
+# silence new console spam from SD2
+ldm.modules.attention.print = lambda *args: None
+ldm.modules.diffusionmodules.model.print = lambda *args: None
def apply_optimizations():
undo_optimizations()
ldm.modules.diffusionmodules.model.nonlinearity = silu
- if cmd_opts.opt_split_attention_v1:
+ if cmd_opts.force_enable_xformers or (cmd_opts.xformers and shared.xformers_available and torch.version.cuda and (6, 0) <= torch.cuda.get_device_capability(shared.device) <= (9, 0)):
+ print("Applying xformers cross attention optimization.")
+ ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.xformers_attention_forward
+ ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.xformers_attnblock_forward
+ elif cmd_opts.opt_split_attention_v1:
+ print("Applying v1 cross attention optimization.")
ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_v1
+ elif not cmd_opts.disable_opt_split_attention and (cmd_opts.opt_split_attention_invokeai or not torch.cuda.is_available()):
+ if not invokeAI_mps_available and shared.device.type == 'mps':
+ print("The InvokeAI cross attention optimization for MPS requires the psutil package which is not installed.")
+ print("Applying v1 cross attention optimization.")
+ ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_v1
+ else:
+ print("Applying cross attention optimization (InvokeAI).")
+ ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_invokeAI
elif not cmd_opts.disable_opt_split_attention and (cmd_opts.opt_split_attention or torch.cuda.is_available()):
+ print("Applying cross attention optimization (Doggettx).")
ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward
ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.cross_attention_attnblock_forward
@@ -37,6 +66,11 @@ def undo_optimizations(): ldm.modules.diffusionmodules.model.AttnBlock.forward = diffusionmodules_model_AttnBlock_forward
+def fix_checkpoint():
+ ldm.modules.attention.BasicTransformerBlock.forward = sd_hijack_checkpoint.BasicTransformerBlock_forward
+ ldm.modules.diffusionmodules.openaimodel.ResBlock.forward = sd_hijack_checkpoint.ResBlock_forward
+ ldm.modules.diffusionmodules.openaimodel.AttentionBlock.forward = sd_hijack_checkpoint.AttentionBlock_forward
+
class StableDiffusionModelHijack:
fixes = None
comments = []
@@ -47,14 +81,18 @@ class StableDiffusionModelHijack: embedding_db = modules.textual_inversion.textual_inversion.EmbeddingDatabase(cmd_opts.embeddings_dir)
def hijack(self, m):
- model_embeddings = m.cond_stage_model.transformer.text_model.embeddings
-
- model_embeddings.token_embedding = EmbeddingsWithFixes(model_embeddings.token_embedding, self)
- m.cond_stage_model = FrozenCLIPEmbedderWithCustomWords(m.cond_stage_model, self)
+ if type(m.cond_stage_model) == ldm.modules.encoders.modules.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(m.cond_stage_model, self)
+ elif type(m.cond_stage_model) == ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder:
+ m.cond_stage_model.model.token_embedding = EmbeddingsWithFixes(m.cond_stage_model.model.token_embedding, self)
+ m.cond_stage_model = sd_hijack_open_clip.FrozenOpenCLIPEmbedderWithCustomWords(m.cond_stage_model, self)
self.clip = m.cond_stage_model
apply_optimizations()
+ fix_checkpoint()
def flatten(el):
flattened = [flatten(children) for children in el.children()]
@@ -66,12 +104,19 @@ class StableDiffusionModelHijack: self.layers = flatten(m)
def undo_hijack(self, m):
- if type(m.cond_stage_model) == FrozenCLIPEmbedderWithCustomWords:
+ if type(m.cond_stage_model) == sd_hijack_clip.FrozenCLIPEmbedderWithCustomWords:
m.cond_stage_model = m.cond_stage_model.wrapped
- model_embeddings = m.cond_stage_model.transformer.text_model.embeddings
- if type(model_embeddings.token_embedding) == EmbeddingsWithFixes:
- model_embeddings.token_embedding = model_embeddings.token_embedding.wrapped
+ model_embeddings = m.cond_stage_model.transformer.text_model.embeddings
+ if type(model_embeddings.token_embedding) == EmbeddingsWithFixes:
+ model_embeddings.token_embedding = model_embeddings.token_embedding.wrapped
+ elif type(m.cond_stage_model) == sd_hijack_open_clip.FrozenOpenCLIPEmbedderWithCustomWords:
+ m.cond_stage_model.wrapped.model.token_embedding = m.cond_stage_model.wrapped.model.token_embedding.wrapped
+ m.cond_stage_model = m.cond_stage_model.wrapped
+
+ self.apply_circular(False)
+ self.layers = None
+ self.clip = None
def apply_circular(self, enable):
if self.circular_enabled == enable:
@@ -82,204 +127,13 @@ class StableDiffusionModelHijack: for layer in [layer for layer in self.layers if type(layer) == torch.nn.Conv2d]:
layer.padding_mode = 'circular' if enable else 'zeros'
+ def clear_comments(self):
+ self.comments = []
+
def tokenize(self, text):
- max_length = self.clip.max_length - 2
_, remade_batch_tokens, _, _, _, token_count = self.clip.process_text([text])
- return remade_batch_tokens[0], token_count, max_length
-
-
-class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
- def __init__(self, wrapped, hijack):
- super().__init__()
- self.wrapped = wrapped
- self.hijack: StableDiffusionModelHijack = hijack
- self.tokenizer = wrapped.tokenizer
- self.max_length = wrapped.max_length
- self.token_mults = {}
-
- tokens_with_parens = [(k, v) for k, v in self.tokenizer.get_vocab().items() if '(' in k or ')' in k or '[' in k or ']' in k]
- for text, ident in tokens_with_parens:
- mult = 1.0
- for c in text:
- if c == '[':
- mult /= 1.1
- if c == ']':
- mult *= 1.1
- if c == '(':
- mult *= 1.1
- if c == ')':
- mult /= 1.1
-
- if mult != 1.0:
- self.token_mults[ident] = mult
-
- def tokenize_line(self, line, used_custom_terms, hijack_comments):
- id_start = self.wrapped.tokenizer.bos_token_id
- id_end = self.wrapped.tokenizer.eos_token_id
- maxlen = self.wrapped.max_length
-
- if opts.enable_emphasis:
- parsed = prompt_parser.parse_prompt_attention(line)
- else:
- parsed = [[line, 1.0]]
-
- tokenized = self.wrapped.tokenizer([text for text, _ in parsed], truncation=False, add_special_tokens=False)["input_ids"]
-
- fixes = []
- remade_tokens = []
- multipliers = []
-
- for tokens, (text, weight) in zip(tokenized, parsed):
- i = 0
- while i < len(tokens):
- token = tokens[i]
-
- embedding, embedding_length_in_tokens = self.hijack.embedding_db.find_embedding_at_position(tokens, i)
-
- if embedding is None:
- remade_tokens.append(token)
- multipliers.append(weight)
- i += 1
- else:
- emb_len = int(embedding.vec.shape[0])
- fixes.append((len(remade_tokens), embedding))
- remade_tokens += [0] * emb_len
- multipliers += [weight] * emb_len
- used_custom_terms.append((embedding.name, embedding.checksum()))
- i += embedding_length_in_tokens
-
- if len(remade_tokens) > maxlen - 2:
- vocab = {v: k for k, v in self.wrapped.tokenizer.get_vocab().items()}
- ovf = remade_tokens[maxlen - 2:]
- overflowing_words = [vocab.get(int(x), "") for x in ovf]
- overflowing_text = self.wrapped.tokenizer.convert_tokens_to_string(''.join(overflowing_words))
- hijack_comments.append(f"Warning: too many input tokens; some ({len(overflowing_words)}) have been truncated:\n{overflowing_text}\n")
-
- token_count = len(remade_tokens)
- remade_tokens = remade_tokens + [id_end] * (maxlen - 2 - len(remade_tokens))
- remade_tokens = [id_start] + remade_tokens[0:maxlen - 2] + [id_end]
-
- multipliers = multipliers + [1.0] * (maxlen - 2 - len(multipliers))
- multipliers = [1.0] + multipliers[0:maxlen - 2] + [1.0]
-
- return remade_tokens, fixes, multipliers, token_count
-
- def process_text(self, texts):
- used_custom_terms = []
- remade_batch_tokens = []
- hijack_comments = []
- hijack_fixes = []
- token_count = 0
-
- cache = {}
- batch_multipliers = []
- for line in texts:
- if line in cache:
- remade_tokens, fixes, multipliers = cache[line]
- else:
- remade_tokens, fixes, multipliers, token_count = self.tokenize_line(line, used_custom_terms, hijack_comments)
-
- cache[line] = (remade_tokens, fixes, multipliers)
-
- remade_batch_tokens.append(remade_tokens)
- hijack_fixes.append(fixes)
- batch_multipliers.append(multipliers)
-
- return batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count
-
+ return remade_batch_tokens[0], token_count, sd_hijack_clip.get_target_prompt_token_count(token_count)
- def process_text_old(self, text):
- id_start = self.wrapped.tokenizer.bos_token_id
- id_end = self.wrapped.tokenizer.eos_token_id
- maxlen = self.wrapped.max_length
- used_custom_terms = []
- remade_batch_tokens = []
- overflowing_words = []
- hijack_comments = []
- hijack_fixes = []
- token_count = 0
-
- cache = {}
- batch_tokens = self.wrapped.tokenizer(text, truncation=False, add_special_tokens=False)["input_ids"]
- batch_multipliers = []
- for tokens in batch_tokens:
- tuple_tokens = tuple(tokens)
-
- if tuple_tokens in cache:
- remade_tokens, fixes, multipliers = cache[tuple_tokens]
- else:
- fixes = []
- remade_tokens = []
- multipliers = []
- mult = 1.0
-
- i = 0
- while i < len(tokens):
- token = tokens[i]
-
- embedding, embedding_length_in_tokens = self.hijack.embedding_db.find_embedding_at_position(tokens, i)
-
- mult_change = self.token_mults.get(token) if opts.enable_emphasis else None
- if mult_change is not None:
- mult *= mult_change
- i += 1
- elif embedding is None:
- remade_tokens.append(token)
- multipliers.append(mult)
- i += 1
- else:
- emb_len = int(embedding.vec.shape[0])
- fixes.append((len(remade_tokens), embedding))
- remade_tokens += [0] * emb_len
- multipliers += [mult] * emb_len
- used_custom_terms.append((embedding.name, embedding.checksum()))
- i += embedding_length_in_tokens
-
- if len(remade_tokens) > maxlen - 2:
- vocab = {v: k for k, v in self.wrapped.tokenizer.get_vocab().items()}
- ovf = remade_tokens[maxlen - 2:]
- overflowing_words = [vocab.get(int(x), "") for x in ovf]
- overflowing_text = self.wrapped.tokenizer.convert_tokens_to_string(''.join(overflowing_words))
- hijack_comments.append(f"Warning: too many input tokens; some ({len(overflowing_words)}) have been truncated:\n{overflowing_text}\n")
-
- token_count = len(remade_tokens)
- remade_tokens = remade_tokens + [id_end] * (maxlen - 2 - len(remade_tokens))
- remade_tokens = [id_start] + remade_tokens[0:maxlen-2] + [id_end]
- cache[tuple_tokens] = (remade_tokens, fixes, multipliers)
-
- multipliers = multipliers + [1.0] * (maxlen - 2 - len(multipliers))
- multipliers = [1.0] + multipliers[0:maxlen - 2] + [1.0]
-
- remade_batch_tokens.append(remade_tokens)
- hijack_fixes.append(fixes)
- batch_multipliers.append(multipliers)
- return batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count
-
- def forward(self, text):
-
- if opts.use_old_emphasis_implementation:
- batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count = self.process_text_old(text)
- else:
- batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count = self.process_text(text)
-
- self.hijack.fixes = hijack_fixes
- self.hijack.comments = hijack_comments
-
- if len(used_custom_terms) > 0:
- self.hijack.comments.append("Used embeddings: " + ", ".join([f'{word} [{checksum}]' for word, checksum in used_custom_terms]))
-
- tokens = torch.asarray(remade_batch_tokens).to(device)
- outputs = self.wrapped.transformer(input_ids=tokens)
- z = outputs.last_hidden_state
-
- # 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(device)
- original_mean = z.mean()
- z *= batch_multipliers.reshape(batch_multipliers.shape + (1,)).expand(z.shape)
- new_mean = z.mean()
- z *= original_mean / new_mean
-
- return z
class EmbeddingsWithFixes(torch.nn.Module):
@@ -301,8 +155,8 @@ class EmbeddingsWithFixes(torch.nn.Module): for fixes, tensor in zip(batch_fixes, inputs_embeds):
for offset, embedding in fixes:
emb = embedding.vec
- emb_len = min(tensor.shape[0]-offset-1, emb.shape[0])
- tensor = torch.cat([tensor[0:offset+1], emb[0:emb_len], tensor[offset+1+emb_len:]])
+ emb_len = min(tensor.shape[0] - offset - 1, emb.shape[0])
+ tensor = torch.cat([tensor[0:offset + 1], emb[0:emb_len], tensor[offset + 1 + emb_len:]])
vecs.append(tensor)
@@ -319,3 +173,23 @@ def add_circular_option_to_conv_2d(): model_hijack = StableDiffusionModelHijack()
+
+
+def register_buffer(self, name, attr):
+ """
+ Fix register buffer bug for Mac OS.
+ """
+
+ if type(attr) == torch.Tensor:
+ if attr.device != devices.device:
+
+ if devices.has_mps():
+ attr = attr.to(device="mps", dtype=torch.float32)
+ else:
+ attr = attr.to(devices.device)
+
+ setattr(self, name, attr)
+
+
+ldm.models.diffusion.ddim.DDIMSampler.register_buffer = register_buffer
+ldm.models.diffusion.plms.PLMSSampler.register_buffer = register_buffer
|