cast long as float before deriving sinusoidal pos emb

dalle2_pytorch/dalle2_pytorch.py

@@ -704,7 +704,7 @@ class Attention(nn.Module):
         sim = sim - sim.amax(dim = -1, keepdim = True).detach()
         sim = sim * self.pb_relax_alpha
 
-        attn = sim.softmax(dim = -1, dtype = torch.float32)
+        attn = sim.softmax(dim = -1)
         attn = self.dropout(attn)
 
         # aggregate values
@@ -1130,7 +1130,7 @@ class SinusoidalPosEmb(nn.Module):
         half_dim = self.dim // 2
         emb = math.log(10000) / (half_dim - 1)
         emb = torch.exp(torch.arange(half_dim, device = x.device) * -emb)
-        emb = rearrange(x, 'i -> i 1') * rearrange(emb, 'j -> 1 j')
+        emb = rearrange(x.type_as(emb), 'i -> i 1') * rearrange(emb, 'j -> 1 j')
         return torch.cat((emb.sin(), emb.cos()), dim = -1)
 
 class Block(nn.Module):
@@ -1272,7 +1272,7 @@ class CrossAttention(nn.Module):
         sim = sim - sim.amax(dim = -1, keepdim = True).detach()
         sim = sim * self.pb_relax_alpha
 
-        attn = sim.softmax(dim = -1, dtype = torch.float32)
+        attn = sim.softmax(dim = -1)
 
         out = einsum('b h i j, b h j d -> b h i d', attn, v)
         out = rearrange(out, 'b h n d -> b n (h d)')

dalle2_pytorch/trainer.py

@@ -550,7 +550,7 @@ class DecoderTrainer(nn.Module):
         if only_model:
             return loaded_obj
 
-        for ind, last_step in zip(range(0, self.num_unets), self.steps.cpu().unbind()):
+        for ind, last_step in zip(range(0, self.num_unets), self.steps.tolist()):
 
             optimizer_key = f'optim{ind}'
             optimizer = getattr(self, optimizer_key)

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '0.16.5'
+__version__ = '0.16.8'