Added lower memory attention from basujindal/stable-diffusion#103

ldm/modules/attention.py

@@ -89,7 +89,7 @@ def forward(self, x):
         b, c, h, w = x.shape
         qkv = self.to_qkv(x)
         q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
-        k = k.softmax(dim=-1)  
+        k = k.softmax(dim=-1)
         context = torch.einsum('bhdn,bhen->bhde', k, v)
         out = torch.einsum('bhde,bhdn->bhen', context, q)
         out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
@@ -174,23 +174,27 @@ def forward(self, x, context=None, mask=None):
         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))
 
-        sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
+        sim = einsum('b i d, b j d -> b i j', q, k) * self.scale  # (8, 4096, 40)
+        del q, k
 
         if exists(mask):
             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)
+            del mask
 
-        # attention, what we cannot get enough of
-        attn = sim.softmax(dim=-1)
+        # attention, what we cannot get enough of, by halves
+        sim[4:] = sim[4:].softmax(dim=-1)
+        sim[:4] = sim[:4].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)
+        sim = einsum('b i j, b j d -> b i d', sim, v)
+        sim = rearrange(sim, '(b h) n d -> b n (h d)', h=h)
+        return self.to_out(sim)
 
 
 class BasicTransformerBlock(nn.Module):
@@ -258,4 +262,4 @@ def forward(self, x, context=None):
             x = block(x, context=context)
         x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w)
         x = self.proj_out(x)
-        return x + x_in
+        return x + x_in