rescale values to prevent linear attention from overflowing in fp16 setting

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -211,6 +211,8 @@ class LinearAttention(nn.Module):
         k = k.softmax(dim = -1)
 
         q = q * self.scale
+        v = v / (h * w)
+
         context = torch.einsum('b h d n, b h e n -> b h d e', k, v)
 
         out = torch.einsum('b h d e, b h d n -> b h e n', context, q)

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.26.3',
+  version = '0.26.4',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',