From 33d7e89c42e0fe7b4a277d7a5bae12ba14828dd8 Mon Sep 17 00:00:00 2001
From: Lime-Cakes <91322985+Lime-Cakes@users.noreply.github.com>
Date: Mon, 14 Nov 2022 20:35:47 +0800
Subject: [PATCH] Edited attention.py for older xformers (#1270)

Older versions of xformers require query, key, value to be contiguous, this calls .contiguous() on q/k/v before passing to xformers.
---
 src/diffusers/models/attention.py | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/src/diffusers/models/attention.py b/src/diffusers/models/attention.py
index e8ea37970e..be9203b4d6 100644
--- a/src/diffusers/models/attention.py
+++ b/src/diffusers/models/attention.py
@@ -557,6 +557,9 @@ class CrossAttention(nn.Module):
         return hidden_states
 
     def _memory_efficient_attention_xformers(self, query, key, value):
+        query = query.contiguous()
+        key = key.contiguous()
+        value = value.contiguous()
         hidden_states = xformers.ops.memory_efficient_attention(query, key, value, attn_bias=None)
         hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
         return hidden_states