From 2eb911b056ce6ff4434f673366782ed34f2b2f12 Mon Sep 17 00:00:00 2001
From: C43H66N12O12S2 <36072735+C43H66N12O12S2@users.noreply.github.com>
Date: Fri, 7 Oct 2022 05:22:28 +0300
Subject: Update sd_hijack.py

---
 modules/sd_hijack.py | 13 +++++++++----
 1 file changed, 9 insertions(+), 4 deletions(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index a6fa890c..6221ed5a 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -20,12 +20,17 @@ diffusionmodules_model_AttnBlock_forward = ldm.modules.diffusionmodules.model.At
 
 
 def apply_optimizations():
-    ldm.modules.diffusionmodules.model.nonlinearity = silu
-
     if cmd_opts.opt_split_attention_v1:
         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 or torch.cuda.is_available()):
-        ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward
+    if cmd_opts.opt_split_attention:
+        ldm.modules.attention_CrossAttention_forward = sd_hijack_optimizations.split_cross_attention_forward
+        ldm.modules.diffusionmodules.model.nonlinearity = sd_hijack_optimizations.nonlinearity_hijack
+        ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.cross_attention_attnblock_forward
+    elif not cmd_opts.disable_opt_xformers_attention:
+        ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.xformers_attention_forward
+        ldm.modules.attention.CrossAttention._maybe_init = sd_hijack_optimizations._maybe_init
+        ldm.modules.attention.CrossAttention.attention_op = None
+        ldm.modules.diffusionmodules.model.nonlinearity = sd_hijack_optimizations.nonlinearity_hijack
         ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.cross_attention_attnblock_forward
 
 
-- 
cgit v1.2.3


From 35d6b231628d18d53d166c3a92fea1523e88d51e Mon Sep 17 00:00:00 2001
From: C43H66N12O12S2 <36072735+C43H66N12O12S2@users.noreply.github.com>
Date: Fri, 7 Oct 2022 05:31:53 +0300
Subject: Update sd_hijack.py

---
 modules/sd_hijack.py | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index 6221ed5a..a006c0a3 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -20,17 +20,16 @@ diffusionmodules_model_AttnBlock_forward = ldm.modules.diffusionmodules.model.At
 
 
 def apply_optimizations():
+    ldm.modules.diffusionmodules.model.nonlinearity = silu
     if cmd_opts.opt_split_attention_v1:
         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_v1
     if cmd_opts.opt_split_attention:
         ldm.modules.attention_CrossAttention_forward = sd_hijack_optimizations.split_cross_attention_forward
-        ldm.modules.diffusionmodules.model.nonlinearity = sd_hijack_optimizations.nonlinearity_hijack
         ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.cross_attention_attnblock_forward
     elif not cmd_opts.disable_opt_xformers_attention:
         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.xformers_attention_forward
         ldm.modules.attention.CrossAttention._maybe_init = sd_hijack_optimizations._maybe_init
         ldm.modules.attention.CrossAttention.attention_op = None
-        ldm.modules.diffusionmodules.model.nonlinearity = sd_hijack_optimizations.nonlinearity_hijack
         ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.cross_attention_attnblock_forward
 
 
-- 
cgit v1.2.3


From 5303df24282ba06abb34a423f2967354d37d078e Mon Sep 17 00:00:00 2001
From: C43H66N12O12S2 <36072735+C43H66N12O12S2@users.noreply.github.com>
Date: Fri, 7 Oct 2022 06:01:14 +0300
Subject: Update sd_hijack.py

---
 modules/sd_hijack.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index a006c0a3..ddacb0ad 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -23,10 +23,10 @@ def apply_optimizations():
     ldm.modules.diffusionmodules.model.nonlinearity = silu
     if cmd_opts.opt_split_attention_v1:
         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_v1
-    if cmd_opts.opt_split_attention:
+    elif cmd_opts.opt_split_attention:
         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
-    elif not cmd_opts.disable_opt_xformers_attention:
+    elif not cmd_opts.disable_opt_xformers_attention and not cmd_opts.opt_split_attention:
         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.xformers_attention_forward
         ldm.modules.attention.CrossAttention._maybe_init = sd_hijack_optimizations._maybe_init
         ldm.modules.attention.CrossAttention.attention_op = None
-- 
cgit v1.2.3


From 5e3ff846c56dc8e1d5c76ea04a8f2f74d7da07fc Mon Sep 17 00:00:00 2001
From: C43H66N12O12S2 <36072735+C43H66N12O12S2@users.noreply.github.com>
Date: Fri, 7 Oct 2022 06:38:01 +0300
Subject: Update sd_hijack.py

---
 modules/sd_hijack.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index ddacb0ad..cbdb9d3c 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -26,7 +26,7 @@ def apply_optimizations():
     elif cmd_opts.opt_split_attention:
         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
-    elif not cmd_opts.disable_opt_xformers_attention and not cmd_opts.opt_split_attention:
+    elif not cmd_opts.disable_opt_xformers_attention and not (cmd_opts.opt_split_attention or torch.version.hip):
         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.xformers_attention_forward
         ldm.modules.attention.CrossAttention._maybe_init = sd_hijack_optimizations._maybe_init
         ldm.modules.attention.CrossAttention.attention_op = None
-- 
cgit v1.2.3


From f7c787eb7c295c27439f4fbdf78c26b8389560be Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Fri, 7 Oct 2022 16:39:51 +0300
Subject: make it possible to use hypernetworks without opt split attention

---
 modules/hypernetwork.py | 42 ++++++++++++++++++++++++++++++++++--------
 modules/sd_hijack.py    |  6 ++++--
 2 files changed, 38 insertions(+), 10 deletions(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/hypernetwork.py b/modules/hypernetwork.py
index c5cf4afa..c7b86682 100644
--- a/modules/hypernetwork.py
+++ b/modules/hypernetwork.py
@@ -4,7 +4,12 @@ import sys
 import traceback
 
 import torch
-from modules import devices
+
+from ldm.util import default
+from modules import devices, shared
+import torch
+from torch import einsum
+from einops import rearrange, repeat
 
 
 class HypernetworkModule(torch.nn.Module):
@@ -48,15 +53,36 @@ def load_hypernetworks(path):
 
     return res
 
-def apply(self, x, context=None, mask=None, original=None):
 
+def attention_CrossAttention_forward(self, x, context=None, mask=None):
+    h = self.heads
+
+    q = self.to_q(x)
+    context = default(context, x)
 
-    if CrossAttention.hypernetwork is not None and context.shape[2] in CrossAttention.hypernetwork:
-        if context.shape[1] == 77 and CrossAttention.noise_cond:
-            context = context + (torch.randn_like(context) * 0.1)
-        h_k, h_v = CrossAttention.hypernetwork[context.shape[2]]
-        k = self.to_k(h_k(context))
-        v = self.to_v(h_v(context))
+    hypernetwork = shared.selected_hypernetwork()
+    hypernetwork_layers = (hypernetwork.layers if hypernetwork is not None else {}).get(context.shape[2], None)
+
+    if hypernetwork_layers is not None:
+        k = self.to_k(hypernetwork_layers[0](context))
+        v = self.to_v(hypernetwork_layers[1](context))
     else:
         k = self.to_k(context)
         v = self.to_v(context)
+
+    q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
+
+    sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
+
+    if mask is not None:
+        mask = rearrange(mask, 'b ... -> b (...)')
+        max_neg_value = -torch.finfo(sim.dtype).max
+        mask = repeat(mask, 'b j -> (b h) () j', h=h)
+        sim.masked_fill_(~mask, max_neg_value)
+
+    # attention, what we cannot get enough of
+    attn = sim.softmax(dim=-1)
+
+    out = einsum('b i j, b j d -> b i d', attn, v)
+    out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
+    return self.to_out(out)
diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index a6fa890c..d68f89cc 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -8,7 +8,7 @@ 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
+from modules import prompt_parser, devices, sd_hijack_optimizations, shared, hypernetwork
 from modules.shared import opts, device, cmd_opts
 
 import ldm.modules.attention
@@ -20,6 +20,8 @@ diffusionmodules_model_AttnBlock_forward = ldm.modules.diffusionmodules.model.At
 
 
 def apply_optimizations():
+    undo_optimizations()
+
     ldm.modules.diffusionmodules.model.nonlinearity = silu
 
     if cmd_opts.opt_split_attention_v1:
@@ -30,7 +32,7 @@ def apply_optimizations():
 
 
 def undo_optimizations():
-    ldm.modules.attention.CrossAttention.forward = attention_CrossAttention_forward
+    ldm.modules.attention.CrossAttention.forward = hypernetwork.attention_CrossAttention_forward
     ldm.modules.diffusionmodules.model.nonlinearity = diffusionmodules_model_nonlinearity
     ldm.modules.diffusionmodules.model.AttnBlock.forward = diffusionmodules_model_AttnBlock_forward
 
-- 
cgit v1.2.3


From b70eaeb2005a5a9593119e7fd32b8072c2a208d5 Mon Sep 17 00:00:00 2001
From: C43H66N12O12S2 <36072735+C43H66N12O12S2@users.noreply.github.com>
Date: Sat, 8 Oct 2022 04:10:35 +0300
Subject: delete broken and unnecessary aliases

---
 modules/sd_hijack.py | 10 ++++------
 1 file changed, 4 insertions(+), 6 deletions(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index cbdb9d3c..0e99c319 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -21,16 +21,14 @@ diffusionmodules_model_AttnBlock_forward = ldm.modules.diffusionmodules.model.At
 
 def apply_optimizations():
     ldm.modules.diffusionmodules.model.nonlinearity = silu
-    if cmd_opts.opt_split_attention_v1:
+    if not cmd_opts.disable_opt_xformers_attention and not (cmd_opts.opt_split_attention or torch.version.hip):
+        ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.xformers_attention_forward
+        ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.cross_attention_attnblock_forward
+    elif cmd_opts.opt_split_attention_v1:
         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_v1
     elif cmd_opts.opt_split_attention:
         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
-    elif not cmd_opts.disable_opt_xformers_attention and not (cmd_opts.opt_split_attention or torch.version.hip):
-        ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.xformers_attention_forward
-        ldm.modules.attention.CrossAttention._maybe_init = sd_hijack_optimizations._maybe_init
-        ldm.modules.attention.CrossAttention.attention_op = None
-        ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.cross_attention_attnblock_forward
 
 
 def undo_optimizations():
-- 
cgit v1.2.3


From 91d66f5520df416db718103d460550ad495e952d Mon Sep 17 00:00:00 2001
From: C43H66N12O12S2 <36072735+C43H66N12O12S2@users.noreply.github.com>
Date: Sat, 8 Oct 2022 11:56:01 +0300
Subject: use new attnblock for xformers path

---
 modules/sd_hijack.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index 0e99c319..3da8c8ce 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -23,7 +23,7 @@ def apply_optimizations():
     ldm.modules.diffusionmodules.model.nonlinearity = silu
     if not cmd_opts.disable_opt_xformers_attention and not (cmd_opts.opt_split_attention or torch.version.hip):
         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.xformers_attention_forward
-        ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.cross_attention_attnblock_forward
+        ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.xformers_attnblock_forward
     elif cmd_opts.opt_split_attention_v1:
         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_v1
     elif cmd_opts.opt_split_attention:
-- 
cgit v1.2.3


From 706d5944a075a6523ea7f00165d630efc085ca22 Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Sat, 8 Oct 2022 13:38:57 +0300
Subject: let user choose his own prompt token count limit

---
 modules/processing.py |  6 ++++++
 modules/sd_hijack.py  | 13 +++++++------
 modules/shared.py     |  5 +++--
 3 files changed, 16 insertions(+), 8 deletions(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/processing.py b/modules/processing.py
index f773a30e..d814d5ac 100644
--- a/modules/processing.py
+++ b/modules/processing.py
@@ -123,6 +123,7 @@ class Processed:
         self.index_of_first_image = index_of_first_image
         self.styles = p.styles
         self.job_timestamp = state.job_timestamp
+        self.max_prompt_tokens = opts.max_prompt_tokens
 
         self.eta = p.eta
         self.ddim_discretize = p.ddim_discretize
@@ -141,6 +142,7 @@ class Processed:
         self.all_subseeds = all_subseeds or [self.subseed]
         self.infotexts = infotexts or [info]
 
+
     def js(self):
         obj = {
             "prompt": self.prompt,
@@ -169,6 +171,7 @@ class Processed:
             "infotexts": self.infotexts,
             "styles": self.styles,
             "job_timestamp": self.job_timestamp,
+            "max_prompt_tokens": self.max_prompt_tokens,
         }
 
         return json.dumps(obj)
@@ -266,6 +269,8 @@ def fix_seed(p):
 def create_infotext(p, all_prompts, all_seeds, all_subseeds, comments, iteration=0, position_in_batch=0):
     index = position_in_batch + iteration * p.batch_size
 
+    max_tokens = getattr(p, 'max_prompt_tokens', opts.max_prompt_tokens)
+
     generation_params = {
         "Steps": p.steps,
         "Sampler": sd_samplers.samplers[p.sampler_index].name,
@@ -281,6 +286,7 @@ def create_infotext(p, all_prompts, all_seeds, all_subseeds, comments, iteration
         "Seed resize from": (None if p.seed_resize_from_w == 0 or p.seed_resize_from_h == 0 else f"{p.seed_resize_from_w}x{p.seed_resize_from_h}"),
         "Denoising strength": getattr(p, 'denoising_strength', None),
         "Eta": (None if p.sampler is None or p.sampler.eta == p.sampler.default_eta else p.sampler.eta),
+        "Max tokens": (None if max_tokens == shared.vanilla_max_prompt_tokens else max_tokens)
     }
 
     generation_params.update(p.extra_generation_params)
diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index d68f89cc..340329c0 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -18,7 +18,6 @@ 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
 
-
 def apply_optimizations():
     undo_optimizations()
 
@@ -83,7 +82,7 @@ class StableDiffusionModelHijack:
             layer.padding_mode = 'circular' if enable else 'zeros'
 
     def tokenize(self, text):
-        max_length = self.clip.max_length - 2
+        max_length = opts.max_prompt_tokens - 2
         _, remade_batch_tokens, _, _, _, token_count = self.clip.process_text([text])
         return remade_batch_tokens[0], token_count, max_length
 
@@ -94,7 +93,6 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
         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]
@@ -116,7 +114,7 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
     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
+        maxlen = opts.max_prompt_tokens
 
         if opts.enable_emphasis:
             parsed = prompt_parser.parse_prompt_attention(line)
@@ -191,7 +189,7 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
     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
+        maxlen = self.wrapped.max_length  # you get to stay at 77
         used_custom_terms = []
         remade_batch_tokens = []
         overflowing_words = []
@@ -268,8 +266,11 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
         if len(used_custom_terms) > 0:
             self.hijack.comments.append("Used embeddings: " + ", ".join([f'{word} [{checksum}]' for word, checksum in used_custom_terms]))
 
+        position_ids_array = [min(x, 75) for x in range(len(remade_batch_tokens[0])-1)] + [76]
+        position_ids = torch.asarray(position_ids_array, device=devices.device).expand((1, -1))
+
         tokens = torch.asarray(remade_batch_tokens).to(device)
-        outputs = self.wrapped.transformer(input_ids=tokens)
+        outputs = self.wrapped.transformer(input_ids=tokens, position_ids=position_ids)
         z = outputs.last_hidden_state
 
         # restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise
diff --git a/modules/shared.py b/modules/shared.py
index 879d8424..864e772c 100644
--- a/modules/shared.py
+++ b/modules/shared.py
@@ -118,8 +118,8 @@ prompt_styles = modules.styles.StyleDatabase(styles_filename)
 interrogator = modules.interrogate.InterrogateModels("interrogate")
 
 face_restorers = []
-# This was moved to webui.py with the other model "setup" calls.
-# modules.sd_models.list_models()
+
+vanilla_max_prompt_tokens = 77
 
 
 def realesrgan_models_names():
@@ -221,6 +221,7 @@ options_templates.update(options_section(('sd', "Stable Diffusion"), {
     "use_old_emphasis_implementation": OptionInfo(False, "Use old emphasis implementation. Can be useful to reproduce old seeds."),
     "enable_batch_seeds": OptionInfo(True, "Make K-diffusion samplers produce same images in a batch as when making a single image"),
     "filter_nsfw": OptionInfo(False, "Filter NSFW content"),
+    "max_prompt_tokens": OptionInfo(vanilla_max_prompt_tokens, f"Max prompt token count. Two tokens are reserved for for start and end. Default is {vanilla_max_prompt_tokens}. Setting this to a different value will result in different pictures for same seed.", gr.Number, {"precision": 0}),
     "random_artist_categories": OptionInfo([], "Allowed categories for random artists selection when using the Roll button", gr.CheckboxGroup, {"choices": artist_db.categories()}),
 }))
 
-- 
cgit v1.2.3


From 4999eb2ef9b30e8c42ca7e4a94d4bbffe4d1f015 Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Sat, 8 Oct 2022 14:25:47 +0300
Subject: do not let user choose his own prompt token count limit

---
 README.md             |  1 +
 modules/processing.py |  5 -----
 modules/sd_hijack.py  | 25 ++++++++++++-------------
 modules/shared.py     |  3 ---
 4 files changed, 13 insertions(+), 21 deletions(-)

(limited to 'modules/sd_hijack.py')

diff --git a/README.md b/README.md
index d6e1d50b..ef9b5e31 100644
--- a/README.md
+++ b/README.md
@@ -65,6 +65,7 @@ Check the [custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-web
 - [Composable-Diffusion](https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/), a way to use multiple prompts at once
      - separate prompts using uppercase `AND`
      - also supports weights for prompts: `a cat :1.2 AND a dog AND a penguin :2.2`
+- No token limit for prompts (original stable diffusion lets you use up to 75 tokens)
 
 ## Installation and Running
 Make sure the required [dependencies](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Dependencies) are met and follow the instructions available for both [NVidia](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs) (recommended) and [AMD](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs) GPUs.
diff --git a/modules/processing.py b/modules/processing.py
index 3657fe69..d5162ddc 100644
--- a/modules/processing.py
+++ b/modules/processing.py
@@ -123,7 +123,6 @@ class Processed:
         self.index_of_first_image = index_of_first_image
         self.styles = p.styles
         self.job_timestamp = state.job_timestamp
-        self.max_prompt_tokens = opts.max_prompt_tokens
 
         self.eta = p.eta
         self.ddim_discretize = p.ddim_discretize
@@ -171,7 +170,6 @@ class Processed:
             "infotexts": self.infotexts,
             "styles": self.styles,
             "job_timestamp": self.job_timestamp,
-            "max_prompt_tokens": self.max_prompt_tokens,
         }
 
         return json.dumps(obj)
@@ -269,8 +267,6 @@ def fix_seed(p):
 def create_infotext(p, all_prompts, all_seeds, all_subseeds, comments, iteration=0, position_in_batch=0):
     index = position_in_batch + iteration * p.batch_size
 
-    max_tokens = getattr(p, 'max_prompt_tokens', opts.max_prompt_tokens)
-
     generation_params = {
         "Steps": p.steps,
         "Sampler": sd_samplers.samplers[p.sampler_index].name,
@@ -286,7 +282,6 @@ def create_infotext(p, all_prompts, all_seeds, all_subseeds, comments, iteration
         "Seed resize from": (None if p.seed_resize_from_w == 0 or p.seed_resize_from_h == 0 else f"{p.seed_resize_from_w}x{p.seed_resize_from_h}"),
         "Denoising strength": getattr(p, 'denoising_strength', None),
         "Eta": (None if p.sampler is None or p.sampler.eta == p.sampler.default_eta else p.sampler.eta),
-        "Max tokens": (None if max_tokens == shared.vanilla_max_prompt_tokens else max_tokens)
     }
 
     generation_params.update(p.extra_generation_params)
diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index 340329c0..2c1332c9 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -36,6 +36,13 @@ def undo_optimizations():
     ldm.modules.diffusionmodules.model.AttnBlock.forward = diffusionmodules_model_AttnBlock_forward
 
 
+def get_target_prompt_token_count(token_count):
+    if token_count < 75:
+        return 75
+
+    return math.ceil(token_count / 10) * 10
+
+
 class StableDiffusionModelHijack:
     fixes = None
     comments = []
@@ -84,7 +91,7 @@ class StableDiffusionModelHijack:
     def tokenize(self, text):
         max_length = opts.max_prompt_tokens - 2
         _, remade_batch_tokens, _, _, _, token_count = self.clip.process_text([text])
-        return remade_batch_tokens[0], token_count, max_length
+        return remade_batch_tokens[0], token_count, get_target_prompt_token_count(token_count)
 
 
 class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
@@ -114,7 +121,6 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
     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 = opts.max_prompt_tokens
 
         if opts.enable_emphasis:
             parsed = prompt_parser.parse_prompt_attention(line)
@@ -146,19 +152,12 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
                     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]
+        prompt_target_length = get_target_prompt_token_count(token_count)
+        tokens_to_add = prompt_target_length - len(remade_tokens) + 1
 
-        multipliers = multipliers + [1.0] * (maxlen - 2 - len(multipliers))
-        multipliers = [1.0] + multipliers[0:maxlen - 2] + [1.0]
+        remade_tokens = [id_start] + remade_tokens + [id_end] * tokens_to_add
+        multipliers = [1.0] + multipliers + [1.0] * tokens_to_add
 
         return remade_tokens, fixes, multipliers, token_count
 
diff --git a/modules/shared.py b/modules/shared.py
index ca462628..475d7e52 100644
--- a/modules/shared.py
+++ b/modules/shared.py
@@ -123,8 +123,6 @@ interrogator = modules.interrogate.InterrogateModels("interrogate")
 
 face_restorers = []
 
-vanilla_max_prompt_tokens = 77
-
 
 def realesrgan_models_names():
     import modules.realesrgan_model
@@ -225,7 +223,6 @@ options_templates.update(options_section(('sd', "Stable Diffusion"), {
     "use_old_emphasis_implementation": OptionInfo(False, "Use old emphasis implementation. Can be useful to reproduce old seeds."),
     "enable_batch_seeds": OptionInfo(True, "Make K-diffusion samplers produce same images in a batch as when making a single image"),
     "filter_nsfw": OptionInfo(False, "Filter NSFW content"),
-    "max_prompt_tokens": OptionInfo(vanilla_max_prompt_tokens, f"Max prompt token count. Two tokens are reserved for for start and end. Default is {vanilla_max_prompt_tokens}. Setting this to a different value will result in different pictures for same seed.", gr.Number, {"precision": 0}),
     "random_artist_categories": OptionInfo([], "Allowed categories for random artists selection when using the Roll button", gr.CheckboxGroup, {"choices": artist_db.categories()}),
 }))
 
-- 
cgit v1.2.3


From 77f4237d1c3af1756e7dab2699e3dcebad5619d6 Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Sat, 8 Oct 2022 15:25:59 +0300
Subject: fix bugs related to variable prompt lengths

---
 modules/sd_hijack.py   | 14 +++++++++-----
 modules/sd_samplers.py | 35 ++++++++++++++++++++++++++++-------
 2 files changed, 37 insertions(+), 12 deletions(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index 2c1332c9..7e7fde0f 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -89,7 +89,6 @@ class StableDiffusionModelHijack:
             layer.padding_mode = 'circular' if enable else 'zeros'
 
     def tokenize(self, text):
-        max_length = opts.max_prompt_tokens - 2
         _, remade_batch_tokens, _, _, _, token_count = self.clip.process_text([text])
         return remade_batch_tokens[0], token_count, get_target_prompt_token_count(token_count)
 
@@ -174,7 +173,8 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
             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)
+                remade_tokens, fixes, multipliers, current_token_count = self.tokenize_line(line, used_custom_terms, hijack_comments)
+                token_count = max(current_token_count, token_count)
 
                 cache[line] = (remade_tokens, fixes, multipliers)
 
@@ -265,15 +265,19 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
         if len(used_custom_terms) > 0:
             self.hijack.comments.append("Used embeddings: " + ", ".join([f'{word} [{checksum}]' for word, checksum in used_custom_terms]))
 
-        position_ids_array = [min(x, 75) for x in range(len(remade_batch_tokens[0])-1)] + [76]
+        target_token_count = get_target_prompt_token_count(token_count) + 2
+
+        position_ids_array = [min(x, 75) for x in range(target_token_count-1)] + [76]
         position_ids = torch.asarray(position_ids_array, device=devices.device).expand((1, -1))
 
-        tokens = torch.asarray(remade_batch_tokens).to(device)
+        remade_batch_tokens_of_same_length = [x + [self.wrapped.tokenizer.eos_token_id] * (target_token_count - len(x)) for x in remade_batch_tokens]
+        tokens = torch.asarray(remade_batch_tokens_of_same_length).to(device)
         outputs = self.wrapped.transformer(input_ids=tokens, position_ids=position_ids)
         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)
+        batch_multipliers_of_same_length = [x + [1.0] * (target_token_count - len(x)) for x in batch_multipliers]
+        batch_multipliers = torch.asarray(batch_multipliers_of_same_length).to(device)
         original_mean = z.mean()
         z *= batch_multipliers.reshape(batch_multipliers.shape + (1,)).expand(z.shape)
         new_mean = z.mean()
diff --git a/modules/sd_samplers.py b/modules/sd_samplers.py
index 13a8b322..eade0dbb 100644
--- a/modules/sd_samplers.py
+++ b/modules/sd_samplers.py
@@ -142,6 +142,16 @@ class VanillaStableDiffusionSampler:
         assert all([len(conds) == 1 for conds in conds_list]), 'composition via AND is not supported for DDIM/PLMS samplers'
         cond = tensor
 
+        # for DDIM, shapes must match, we can't just process cond and uncond independently;
+        # filling unconditional_conditioning with repeats of the last vector to match length is
+        # not 100% correct but should work well enough
+        if unconditional_conditioning.shape[1] < cond.shape[1]:
+            last_vector = unconditional_conditioning[:, -1:]
+            last_vector_repeated = last_vector.repeat([1, cond.shape[1] - unconditional_conditioning.shape[1], 1])
+            unconditional_conditioning = torch.hstack([unconditional_conditioning, last_vector_repeated])
+        elif unconditional_conditioning.shape[1] > cond.shape[1]:
+            unconditional_conditioning = unconditional_conditioning[:, :cond.shape[1]]
+
         if self.mask is not None:
             img_orig = self.sampler.model.q_sample(self.init_latent, ts)
             x_dec = img_orig * self.mask + self.nmask * x_dec
@@ -221,18 +231,29 @@ class CFGDenoiser(torch.nn.Module):
 
         x_in = torch.cat([torch.stack([x[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [x])
         sigma_in = torch.cat([torch.stack([sigma[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [sigma])
-        cond_in = torch.cat([tensor, uncond])
 
-        if shared.batch_cond_uncond:
-            x_out = self.inner_model(x_in, sigma_in, cond=cond_in)
+        if tensor.shape[1] == uncond.shape[1]:
+            cond_in = torch.cat([tensor, uncond])
+
+            if shared.batch_cond_uncond:
+                x_out = self.inner_model(x_in, sigma_in, cond=cond_in)
+            else:
+                x_out = torch.zeros_like(x_in)
+                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=cond_in[a:b])
         else:
             x_out = torch.zeros_like(x_in)
-            for batch_offset in range(0, x_out.shape[0], batch_size):
+            batch_size = batch_size*2 if shared.batch_cond_uncond else batch_size
+            for batch_offset in range(0, tensor.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=cond_in[a:b])
+                b = min(a + batch_size, tensor.shape[0])
+                x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond=tensor[a:b])
+
+            x_out[-uncond.shape[0]:] = self.inner_model(x_in[-uncond.shape[0]:], sigma_in[-uncond.shape[0]:], cond=uncond)
 
-        denoised_uncond = x_out[-batch_size:]
+        denoised_uncond = x_out[-uncond.shape[0]:]
         denoised = torch.clone(denoised_uncond)
 
         for i, conds in enumerate(conds_list):
-- 
cgit v1.2.3


From 5f85a74b00c0154bfd559dc67edfa7e30342b7c9 Mon Sep 17 00:00:00 2001
From: MrCheeze <fishycheeze@yahoo.ca>
Date: Fri, 7 Oct 2022 17:48:34 -0400
Subject: fix bug where when using prompt composition, hijack_comments
 generated before the final AND will be dropped

---
 modules/processing.py | 1 +
 modules/sd_hijack.py  | 5 ++++-
 2 files changed, 5 insertions(+), 1 deletion(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/processing.py b/modules/processing.py
index d5162ddc..8240ee27 100644
--- a/modules/processing.py
+++ b/modules/processing.py
@@ -313,6 +313,7 @@ def process_images(p: StableDiffusionProcessing) -> Processed:
         os.makedirs(p.outpath_grids, exist_ok=True)
 
     modules.sd_hijack.model_hijack.apply_circular(p.tiling)
+    modules.sd_hijack.model_hijack.clear_comments()
 
     comments = {}
 
diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index 7e7fde0f..ba808a39 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -88,6 +88,9 @@ 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):
         _, remade_batch_tokens, _, _, _, token_count = self.clip.process_text([text])
         return remade_batch_tokens[0], token_count, get_target_prompt_token_count(token_count)
@@ -260,7 +263,7 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
             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
+        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]))
-- 
cgit v1.2.3


From 26b459a3799c5cdf71ca8ed5315a99f69c69f02c Mon Sep 17 00:00:00 2001
From: C43H66N12O12S2 <36072735+C43H66N12O12S2@users.noreply.github.com>
Date: Sat, 8 Oct 2022 16:20:04 +0300
Subject: default to split attention if cuda is available and xformers is not

---
 modules/sd_hijack.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index 3da8c8ce..04adcf03 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -21,12 +21,12 @@ diffusionmodules_model_AttnBlock_forward = ldm.modules.diffusionmodules.model.At
 
 def apply_optimizations():
     ldm.modules.diffusionmodules.model.nonlinearity = silu
-    if not cmd_opts.disable_opt_xformers_attention and not (cmd_opts.opt_split_attention or torch.version.hip):
+    if not cmd_opts.disable_opt_xformers_attention and not (cmd_opts.opt_split_attention or torch.version.hip or shared.xformers_available):
         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:
         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_v1
-    elif cmd_opts.opt_split_attention:
+    elif cmd_opts.opt_split_attention or torch.cuda.is_available():
         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
 
-- 
cgit v1.2.3


From 970de9ee6891ff586821d0d80dde01c2f6c681b3 Mon Sep 17 00:00:00 2001
From: C43H66N12O12S2 <36072735+C43H66N12O12S2@users.noreply.github.com>
Date: Sat, 8 Oct 2022 16:29:43 +0300
Subject: Update sd_hijack.py

---
 modules/sd_hijack.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index 04adcf03..5b30539f 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -21,7 +21,7 @@ diffusionmodules_model_AttnBlock_forward = ldm.modules.diffusionmodules.model.At
 
 def apply_optimizations():
     ldm.modules.diffusionmodules.model.nonlinearity = silu
-    if not cmd_opts.disable_opt_xformers_attention and not (cmd_opts.opt_split_attention or torch.version.hip or shared.xformers_available):
+    if not cmd_opts.disable_opt_xformers_attention and not (cmd_opts.opt_split_attention or torch.version.hip) and shared.xformers_available:
         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:
-- 
cgit v1.2.3


From dc1117233ef8f9b25ff1ac40b158f20b70ba2fcb Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Sat, 8 Oct 2022 17:02:18 +0300
Subject: simplify xfrmers options: --xformers to enable and that's it

---
 launch.py                          |  2 +-
 modules/sd_hijack.py               |  2 +-
 modules/sd_hijack_optimizations.py | 20 +++++++++++++-------
 modules/shared.py                  |  2 +-
 4 files changed, 16 insertions(+), 10 deletions(-)

(limited to 'modules/sd_hijack.py')

diff --git a/launch.py b/launch.py
index a592e1ba..61f62096 100644
--- a/launch.py
+++ b/launch.py
@@ -32,7 +32,7 @@ def extract_arg(args, name):
 
 
 args, skip_torch_cuda_test = extract_arg(args, '--skip-torch-cuda-test')
-args, xformers = extract_arg(args, '--xformers')
+xformers = '--xformers' in args
 
 
 def repo_dir(name):
diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index 5d93f7f6..91e98c16 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -22,7 +22,7 @@ def apply_optimizations():
     undo_optimizations()
 
     ldm.modules.diffusionmodules.model.nonlinearity = silu
-    if not cmd_opts.disable_opt_xformers_attention and not (cmd_opts.opt_split_attention or torch.version.hip) and shared.xformers_available:
+    if cmd_opts.xformers and shared.xformers_available and not torch.version.hip:
         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:
diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py
index 05023b6f..d23d733b 100644
--- a/modules/sd_hijack_optimizations.py
+++ b/modules/sd_hijack_optimizations.py
@@ -1,4 +1,7 @@
 import math
+import sys
+import traceback
+
 import torch
 from torch import einsum
 
@@ -7,13 +10,16 @@ from einops import rearrange
 
 from modules import shared
 
-try:
-    import xformers.ops
-    import functorch
-    xformers._is_functorch_available = True
-    shared.xformers_available = True
-except Exception:
-    print('Cannot find xformers, defaulting to split attention. Try adding --xformers commandline argument to your webui-user file if you wish to install it.')
+if shared.cmd_opts.xformers:
+    try:
+        import xformers.ops
+        import functorch
+        xformers._is_functorch_available = True
+        shared.xformers_available = True
+    except Exception:
+        print("Cannot import xformers", file=sys.stderr)
+        print(traceback.format_exc(), file=sys.stderr)
+
 
 # see https://github.com/basujindal/stable-diffusion/pull/117 for discussion
 def split_cross_attention_forward_v1(self, x, context=None, mask=None):
diff --git a/modules/shared.py b/modules/shared.py
index d68df751..02cb2722 100644
--- a/modules/shared.py
+++ b/modules/shared.py
@@ -43,7 +43,7 @@ parser.add_argument("--realesrgan-models-path", type=str, help="Path to director
 parser.add_argument("--scunet-models-path", type=str, help="Path to directory with ScuNET model file(s).", default=os.path.join(models_path, 'ScuNET'))
 parser.add_argument("--swinir-models-path", type=str, help="Path to directory with SwinIR model file(s).", default=os.path.join(models_path, 'SwinIR'))
 parser.add_argument("--ldsr-models-path", type=str, help="Path to directory with LDSR model file(s).", default=os.path.join(models_path, 'LDSR'))
-parser.add_argument("--disable-opt-xformers-attention", action='store_true', help="force-disables xformers attention optimization")
+parser.add_argument("--xformers", action='store_true', help="enable xformers for cross attention layers")
 parser.add_argument("--opt-split-attention", action='store_true', help="force-enables cross-attention layer optimization. By default, it's on for torch.cuda and off for other torch devices.")
 parser.add_argument("--disable-opt-split-attention", action='store_true', help="force-disables cross-attention layer optimization")
 parser.add_argument("--opt-split-attention-v1", action='store_true', help="enable older version of split attention optimization that does not consume all the VRAM it can find")
-- 
cgit v1.2.3


From 27032c47df9c07ac21dd5b89fa7dc247bb8705b6 Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Sat, 8 Oct 2022 17:10:05 +0300
Subject: restore old opt_split_attention/disable_opt_split_attention logic

---
 modules/sd_hijack.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index 91e98c16..335a2bcf 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -27,7 +27,7 @@ def apply_optimizations():
         ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.xformers_attnblock_forward
     elif cmd_opts.opt_split_attention_v1:
         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_v1
-    elif cmd_opts.opt_split_attention or torch.cuda.is_available():
+    elif not cmd_opts.disable_opt_split_attention and (cmd_opts.opt_split_attention or torch.cuda.is_available()):
         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
 
-- 
cgit v1.2.3


From cfc33f99d47d1f45af15499e5965834089d11858 Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Sat, 8 Oct 2022 17:28:58 +0300
Subject: why did you do this

---
 modules/sd_hijack.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index 335a2bcf..ed271976 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -28,7 +28,7 @@ def apply_optimizations():
     elif cmd_opts.opt_split_attention_v1:
         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 or torch.cuda.is_available()):
-        ldm.modules.attention_CrossAttention_forward = sd_hijack_optimizations.split_cross_attention_forward
+        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
 
 
-- 
cgit v1.2.3


From 017b6b8744f0771e498656ec043e12d5cc6969a7 Mon Sep 17 00:00:00 2001
From: C43H66N12O12S2 <36072735+C43H66N12O12S2@users.noreply.github.com>
Date: Sat, 8 Oct 2022 17:27:21 +0300
Subject: check for ampere

---
 modules/sd_hijack.py | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index ed271976..5e266d5e 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -22,9 +22,10 @@ def apply_optimizations():
     undo_optimizations()
 
     ldm.modules.diffusionmodules.model.nonlinearity = silu
-    if cmd_opts.xformers and shared.xformers_available and not torch.version.hip:
-        ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.xformers_attention_forward
-        ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.xformers_attnblock_forward
+    if cmd_opts.xformers and shared.xformers_available and torch.version.cuda:
+        if torch.cuda.get_device_capability(shared.device) == (8, 6):
+            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:
         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 or torch.cuda.is_available()):
-- 
cgit v1.2.3


From cc0258aea7b6605be3648900063cfa96ed7c5ffa Mon Sep 17 00:00:00 2001
From: C43H66N12O12S2 <36072735+C43H66N12O12S2@users.noreply.github.com>
Date: Sat, 8 Oct 2022 17:44:53 +0300
Subject: check for ampere without destroying the optimizations. again.

---
 modules/sd_hijack.py | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index 5e266d5e..a3e374f0 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -22,10 +22,9 @@ def apply_optimizations():
     undo_optimizations()
 
     ldm.modules.diffusionmodules.model.nonlinearity = silu
-    if cmd_opts.xformers and shared.xformers_available and torch.version.cuda:
-        if torch.cuda.get_device_capability(shared.device) == (8, 6):
-            ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.xformers_attention_forward
-            ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.xformers_attnblock_forward
+    if cmd_opts.xformers and shared.xformers_available and torch.version.cuda and torch.cuda.get_device_capability(shared.device) == (8, 6):
+        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:
         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 or torch.cuda.is_available()):
-- 
cgit v1.2.3


From 3061cdb7b610d4ba7f1ea695d9d6364b591e5bc7 Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Sat, 8 Oct 2022 19:22:15 +0300
Subject: add --force-enable-xformers option and also add messages to console
 regarding cross attention optimizations

---
 modules/sd_hijack.py | 6 +++++-
 modules/shared.py    | 1 +
 2 files changed, 6 insertions(+), 1 deletion(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index a3e374f0..307cc67d 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -22,12 +22,16 @@ def apply_optimizations():
     undo_optimizations()
 
     ldm.modules.diffusionmodules.model.nonlinearity = silu
-    if cmd_opts.xformers and shared.xformers_available and torch.version.cuda and torch.cuda.get_device_capability(shared.device) == (8, 6):
+
+    if cmd_opts.force_enable_xformers or (cmd_opts.xformers and shared.xformers_available and torch.version.cuda and torch.cuda.get_device_capability(shared.device) == (8, 6)):
+        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 or torch.cuda.is_available()):
+        print("Applying cross attention optimization.")
         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
 
diff --git a/modules/shared.py b/modules/shared.py
index 02cb2722..8f941226 100644
--- a/modules/shared.py
+++ b/modules/shared.py
@@ -44,6 +44,7 @@ parser.add_argument("--scunet-models-path", type=str, help="Path to directory wi
 parser.add_argument("--swinir-models-path", type=str, help="Path to directory with SwinIR model file(s).", default=os.path.join(models_path, 'SwinIR'))
 parser.add_argument("--ldsr-models-path", type=str, help="Path to directory with LDSR model file(s).", default=os.path.join(models_path, 'LDSR'))
 parser.add_argument("--xformers", action='store_true', help="enable xformers for cross attention layers")
+parser.add_argument("--force-enable-xformers", action='store_true', help="enable xformers for cross attention layers regardless of whether the checking code thinks you can run it; do not make bug reports if this fails to work")
 parser.add_argument("--opt-split-attention", action='store_true', help="force-enables cross-attention layer optimization. By default, it's on for torch.cuda and off for other torch devices.")
 parser.add_argument("--disable-opt-split-attention", action='store_true', help="force-disables cross-attention layer optimization")
 parser.add_argument("--opt-split-attention-v1", action='store_true', help="enable older version of split attention optimization that does not consume all the VRAM it can find")
-- 
cgit v1.2.3


From 1371d7608b402d6f15c200ec2f5fde4579836a05 Mon Sep 17 00:00:00 2001
From: Fampai <unknown>
Date: Sat, 8 Oct 2022 14:28:22 -0400
Subject: Added ability to ignore last n layers in FrozenCLIPEmbedder

---
 modules/sd_hijack.py | 11 +++++++++--
 modules/shared.py    |  1 +
 2 files changed, 10 insertions(+), 2 deletions(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index 307cc67d..f12a9696 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -281,8 +281,15 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
 
         remade_batch_tokens_of_same_length = [x + [self.wrapped.tokenizer.eos_token_id] * (target_token_count - len(x)) for x in remade_batch_tokens]
         tokens = torch.asarray(remade_batch_tokens_of_same_length).to(device)
-        outputs = self.wrapped.transformer(input_ids=tokens, position_ids=position_ids)
-        z = outputs.last_hidden_state
+
+        tmp = -opts.CLIP_ignore_last_layers
+        if (opts.CLIP_ignore_last_layers == 0):
+            outputs = self.wrapped.transformer(input_ids=tokens, position_ids=position_ids)
+            z = outputs.last_hidden_state
+        else:
+            outputs = self.wrapped.transformer(input_ids=tokens, position_ids=position_ids, output_hidden_states=tmp)
+            z = outputs.hidden_states[tmp]
+            z = self.wrapped.transformer.text_model.final_layer_norm(z)
 
         # restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise
         batch_multipliers_of_same_length = [x + [1.0] * (target_token_count - len(x)) for x in batch_multipliers]
diff --git a/modules/shared.py b/modules/shared.py
index 8f941226..af8dc744 100644
--- a/modules/shared.py
+++ b/modules/shared.py
@@ -225,6 +225,7 @@ options_templates.update(options_section(('sd', "Stable Diffusion"), {
     "use_old_emphasis_implementation": OptionInfo(False, "Use old emphasis implementation. Can be useful to reproduce old seeds."),
     "enable_batch_seeds": OptionInfo(True, "Make K-diffusion samplers produce same images in a batch as when making a single image"),
     "filter_nsfw": OptionInfo(False, "Filter NSFW content"),
+    'CLIP_ignore_last_layers': OptionInfo(0, "Ignore last layers of CLIP model", gr.Slider, {"minimum": 0, "maximum": 5, "step": 1}),
     "random_artist_categories": OptionInfo([], "Allowed categories for random artists selection when using the Roll button", gr.CheckboxGroup, {"choices": artist_db.categories()}),
 }))
 
-- 
cgit v1.2.3


From e59c66c0088422b27f64b401ef42c242f836725a Mon Sep 17 00:00:00 2001
From: Fampai <unknown>
Date: Sat, 8 Oct 2022 16:32:05 -0400
Subject: Optimized code for Ignoring last CLIP layers

---
 modules/sd_hijack.py | 12 ++++--------
 1 file changed, 4 insertions(+), 8 deletions(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index f12a9696..4a2d2153 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -282,14 +282,10 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
         remade_batch_tokens_of_same_length = [x + [self.wrapped.tokenizer.eos_token_id] * (target_token_count - len(x)) for x in remade_batch_tokens]
         tokens = torch.asarray(remade_batch_tokens_of_same_length).to(device)
 
-        tmp = -opts.CLIP_ignore_last_layers
-        if (opts.CLIP_ignore_last_layers == 0):
-            outputs = self.wrapped.transformer(input_ids=tokens, position_ids=position_ids)
-            z = outputs.last_hidden_state
-        else:
-            outputs = self.wrapped.transformer(input_ids=tokens, position_ids=position_ids, output_hidden_states=tmp)
-            z = outputs.hidden_states[tmp]
-            z = self.wrapped.transformer.text_model.final_layer_norm(z)
+        tmp = -opts.CLIP_stop_at_last_layers
+        outputs = self.wrapped.transformer(input_ids=tokens, position_ids=position_ids, output_hidden_states=tmp)
+        z = outputs.hidden_states[tmp]
+        z = self.wrapped.transformer.text_model.final_layer_norm(z)
 
         # restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise
         batch_multipliers_of_same_length = [x + [1.0] * (target_token_count - len(x)) for x in batch_multipliers]
-- 
cgit v1.2.3


From ad3ae441081155dcd4fde805279e5082ca264695 Mon Sep 17 00:00:00 2001
From: Fampai <unknown>
Date: Sun, 9 Oct 2022 04:32:40 -0400
Subject: Updated code for legibility

---
 modules/sd_hijack.py | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index 4a2d2153..7793d25b 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -284,8 +284,11 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
 
         tmp = -opts.CLIP_stop_at_last_layers
         outputs = self.wrapped.transformer(input_ids=tokens, position_ids=position_ids, output_hidden_states=tmp)
-        z = outputs.hidden_states[tmp]
-        z = self.wrapped.transformer.text_model.final_layer_norm(z)
+        if tmp < -1:
+            z = outputs.hidden_states[tmp]
+            z = self.wrapped.transformer.text_model.final_layer_norm(z)
+        else:
+            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_of_same_length = [x + [1.0] * (target_token_count - len(x)) for x in batch_multipliers]
-- 
cgit v1.2.3


From 1824e9ee3ab4f94aee8908a62ea2569a01aeb3d7 Mon Sep 17 00:00:00 2001
From: Fampai <unknown>
Date: Sun, 9 Oct 2022 14:15:43 -0400
Subject: Removed unnecessary tmp variable

---
 modules/sd_hijack.py | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

(limited to 'modules/sd_hijack.py')

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index 7793d25b..437acce4 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -282,10 +282,9 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
         remade_batch_tokens_of_same_length = [x + [self.wrapped.tokenizer.eos_token_id] * (target_token_count - len(x)) for x in remade_batch_tokens]
         tokens = torch.asarray(remade_batch_tokens_of_same_length).to(device)
 
-        tmp = -opts.CLIP_stop_at_last_layers
-        outputs = self.wrapped.transformer(input_ids=tokens, position_ids=position_ids, output_hidden_states=tmp)
-        if tmp < -1:
-            z = outputs.hidden_states[tmp]
+        outputs = self.wrapped.transformer(input_ids=tokens, position_ids=position_ids, output_hidden_states=-opts.CLIP_stop_at_last_layers)
+        if opts.CLIP_stop_at_last_layers > 1:
+            z = outputs.hidden_states[-opts.CLIP_stop_at_last_layers]
             z = self.wrapped.transformer.text_model.final_layer_norm(z)
         else:
             z = outputs.last_hidden_state
-- 
cgit v1.2.3