From 06fadd2dc5c2753558a9f3971568c2673819f48c Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Sun, 11 Sep 2022 00:29:10 +0300
Subject: [PATCH] added --opt-split-attention-v1

---
 modules/sd_hijack.py | 33 +++++++++++++++++++++++++++++++++
 modules/shared.py    |  1 +
 2 files changed, 34 insertions(+)

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index 60bc6671..9eb6cc20 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -13,6 +13,37 @@ from einops import rearrange
 import ldm.modules.attention
 
 
+
+# see https://github.com/basujindal/stable-diffusion/pull/117 for discussion
+def split_cross_attention_forward_v1(self, x, context=None, mask=None):
+    h = self.heads
+
+    q = self.to_q(x)
+    context = default(context, x)
+    k = self.to_k(context)
+    v = self.to_v(context)
+    del context, x
+
+    q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
+
+    r1 = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device)
+    for i in range(0, q.shape[0], 2):
+        end = i + 2
+        s1 = einsum('b i d, b j d -> b i j', q[i:end], k[i:end])
+        s1 *= self.scale
+
+        s2 = s1.softmax(dim=-1)
+        del s1
+
+        r1[i:end] = einsum('b i j, b j d -> b i d', s2, v[i:end])
+        del s2
+
+    r2 = rearrange(r1, '(b h) n d -> b n (h d)', h=h)
+    del r1
+
+    return self.to_out(r2)
+
+
 # taken from https://github.com/Doggettx/stable-diffusion
 def split_cross_attention_forward(self, x, context=None, mask=None):
     h = self.heads
@@ -143,6 +174,8 @@ class StableDiffusionModelHijack:
 
         if cmd_opts.opt_split_attention:
             ldm.modules.attention.CrossAttention.forward = split_cross_attention_forward
+        elif cmd_opts.opt_split_attention_v1:
+            ldm.modules.attention.CrossAttention.forward = split_cross_attention_forward_v1
 
         def flatten(el):
             flattened = [flatten(children) for children in el.children()]
diff --git a/modules/shared.py b/modules/shared.py
index 614fd5a3..6ca9106c 100644
--- a/modules/shared.py
+++ b/modules/shared.py
@@ -35,6 +35,7 @@ parser.add_argument("--precision", type=str, help="evaluate at this precision",
 parser.add_argument("--share", action='store_true', help="use share=True for gradio and make the UI accessible through their site (doesn't work for me but you might have better luck)")
 parser.add_argument("--esrgan-models-path", type=str, help="path to directory with ESRGAN models", default=os.path.join(script_path, 'ESRGAN'))
 parser.add_argument("--opt-split-attention", action='store_true', help="enable optimization that reduce vram usage by a lot for about 10%% decrease in performance")
+parser.add_argument("--opt-split-attention-v1", action='store_true', help="enable older version of --opt-split-attention optimization")
 parser.add_argument("--listen", action='store_true', help="launch gradio with 0.0.0.0 as server name, allowing to respond to network requests")
 parser.add_argument("--port", type=int, help="launch gradio with given server port, you need root/admin rights for ports < 1024, defaults to 7860 if available", default=None)
 parser.add_argument("--show-negative-prompt", action='store_true', help="does not do anything", default=False)