Enable even larger images with one simple torch.nn.functional.silu im…

…port (invoke-ai#653)

Fixes:
File "stable-diffusion/ldm/modules/diffusionmodules/model.py", line 37, in nonlinearity
    return x*torch.sigmoid(x)
RuntimeError: CUDA out of memory. Tried to allocate 1.56 GiB [..]

Now up to 1536x1280 is possible on 8GB VRAM.
Also remove unused SiLU class.

ldm/modules/diffusionmodules/model.py

@@ -3,6 +3,7 @@
 import math
 import torch
 import torch.nn as nn
+from torch.nn.functional import silu
 import numpy as np
 from einops import rearrange
 
@@ -32,11 +33,6 @@ def get_timestep_embedding(timesteps, embedding_dim):
     return emb
 
 
-def nonlinearity(x):
-    # swish
-    return x*torch.sigmoid(x)
-
-
 def Normalize(in_channels, num_groups=32):
     return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
 
@@ -122,14 +118,14 @@ def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False,
 
     def forward(self, x, temb):
         h = self.norm1(x)
-        h = nonlinearity(h)
+        h = silu(h)
         h = self.conv1(h)
 
         if temb is not None:
-            h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None]
+            h = h + self.temb_proj(silu(temb))[:,:,None,None]
 
         h = self.norm2(h)
-        h = nonlinearity(h)
+        h = silu(h)
         h = self.dropout(h)
         h = self.conv2(h)
 
@@ -368,7 +364,7 @@ def forward(self, x, t=None, context=None):
             assert t is not None
             temb = get_timestep_embedding(t, self.ch)
             temb = self.temb.dense[0](temb)
-            temb = nonlinearity(temb)
+            temb = silu(temb)
             temb = self.temb.dense[1](temb)
         else:
             temb = None
@@ -402,7 +398,7 @@ def forward(self, x, t=None, context=None):
 
         # end
         h = self.norm_out(h)
-        h = nonlinearity(h)
+        h = silu(h)
         h = self.conv_out(h)
         return h
 
@@ -499,7 +495,7 @@ def forward(self, x):
 
         # end
         h = self.norm_out(h)
-        h = nonlinearity(h)
+        h = silu(h)
         h = self.conv_out(h)
         return h
 
@@ -611,7 +607,7 @@ def forward(self, z):
             return h
 
         h = self.norm_out(h)
-        h = nonlinearity(h)
+        h = silu(h)
         h = self.conv_out(h)
         if self.tanh_out:
             h = torch.tanh(h)
@@ -649,7 +645,7 @@ def forward(self, x):
                 x = layer(x)
 
         h = self.norm_out(x)
-        h = nonlinearity(h)
+        h = silu(h)
         x = self.conv_out(h)
         return x
 
@@ -697,7 +693,7 @@ def forward(self, x):
             if i_level != self.num_resolutions - 1:
                 h = self.upsample_blocks[k](h)
         h = self.norm_out(h)
-        h = nonlinearity(h)
+        h = silu(h)
         h = self.conv_out(h)
         return h
 
@@ -873,7 +869,7 @@ def forward(self,x):
         z_fs = self.encode_with_pretrained(x)
         z = self.proj_norm(z_fs)
         z = self.proj(z)
-        z = nonlinearity(z)
+        z = silu(z)
 
         for submodel, downmodel in zip(self.model,self.downsampler):
             z = submodel(z,temb=None)

ldm/modules/diffusionmodules/util.py

@@ -252,12 +252,6 @@ def normalization(channels):
     return GroupNorm32(32, channels)
 
 
-# PyTorch 1.7 has SiLU, but we support PyTorch 1.5.
-class SiLU(nn.Module):
-    def forward(self, x):
-        return x * torch.sigmoid(x)
-
-
 class GroupNorm32(nn.GroupNorm):
     def forward(self, x):
         return super().forward(x.float()).type(x.dtype)