debugging with Aidan

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
@@ -1127,11 +1127,12 @@ class SinusoidalPosEmb(nn.Module):
         self.dim = dim
 
     def forward(self, x):
+        dtype, device = x.dtype, x.device
         half_dim = self.dim // 2
         emb = math.log(10000) / (half_dim - 1)
-        emb = torch.exp(torch.arange(half_dim, device = x.device) * -emb)
+        emb = torch.exp(torch.arange(half_dim, device = device, dtype = dtype) * -emb)
         emb = rearrange(x, 'i -> i 1') * rearrange(emb, 'j -> 1 j')
-        return torch.cat((emb.sin(), emb.cos()), dim = -1)
+        return torch.cat((emb.sin(), emb.cos()), dim = -1).type(dtype)
 
 class Block(nn.Module):
     def __init__(
@@ -1272,7 +1273,7 @@ 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)')
@@ -1626,6 +1627,7 @@ class Unet(nn.Module):
 
         # time conditioning
 
+        time = time.type_as(x)
         time_hiddens = self.to_time_hiddens(time)
 
         time_tokens = self.to_time_tokens(time_hiddens)

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '0.16.6'
+__version__ = '0.16.9'