From 8d7fa2f67cb0554d8902d5d407166876020e067e Mon Sep 17 00:00:00 2001
From: Pam <pamhome21@gmail.com>
Date: Fri, 10 Mar 2023 22:48:41 +0500
Subject: [PATCH] sdp_attnblock_forward hijack

---
 modules/sd_hijack.py               |  2 ++
 modules/sd_hijack_optimizations.py | 24 ++++++++++++++++++++++++
 2 files changed, 26 insertions(+)

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index e98ae51a..f4bb0266 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -47,10 +47,12 @@ def apply_optimizations():
     elif cmd_opts.opt_sdp_no_mem_attention and can_use_sdp:
         print("Applying scaled dot product cross attention optimization (without memory efficient attention).")
         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.scaled_dot_product_no_mem_attention_forward
+        ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.sdp_no_mem_attnblock_forward
         optimization_method = 'sdp-no-mem'
     elif cmd_opts.opt_sdp_attention and can_use_sdp:
         print("Applying scaled dot product cross attention optimization.")
         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.scaled_dot_product_attention_forward
+        ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.sdp_attnblock_forward
         optimization_method = 'sdp'
     elif cmd_opts.opt_sub_quad_attention:
         print("Applying sub-quadratic cross attention optimization.")
diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py
index 68b1dd84..2e307b5d 100644
--- a/modules/sd_hijack_optimizations.py
+++ b/modules/sd_hijack_optimizations.py
@@ -473,6 +473,30 @@ 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_)
+    q = self.q(h_)
+    k = self.k(h_)
+    v = self.v(h_)
+    b, c, h, w = q.shape
+    q, k, v = map(lambda t: rearrange(t, 'b c h w -> b (h w) c'), (q, k, v))
+    dtype = q.dtype
+    if shared.opts.upcast_attn:
+        q, k = q.float(), k.float()
+    q = q.contiguous()
+    k = k.contiguous()
+    v = v.contiguous()
+    out = torch.nn.functional.scaled_dot_product_attention(q, k, v, dropout_p=0.0, is_causal=False)
+    out = out.to(dtype)
+    out = rearrange(out, 'b (h w) c -> b c h w', h=h)
+    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_)