fix a dtype issue when evaluating the sana transformer with a float16 autocast context

Author: Nerogar

src/diffusers/models/attention_processor.py

@@ -5896,9 +5896,10 @@ def __call__(
 
         query, key, value = query.float(), key.float(), value.float()
 
-        value = F.pad(value, (0, 0, 0, 1), mode="constant", value=1.0)
-        scores = torch.matmul(value, key)
-        hidden_states = torch.matmul(scores, query)
+        with torch.autocast(device_type=hidden_states.device.type, enabled=False):
+            value = F.pad(value, (0, 0, 0, 1), mode="constant", value=1.0)
+            scores = torch.matmul(value, key)
+            hidden_states = torch.matmul(scores, query)
 
         hidden_states = hidden_states[:, :, :-1] / (hidden_states[:, :, -1:] + 1e-15)
         hidden_states = hidden_states.flatten(1, 2).transpose(1, 2)