[DC-AE, SANA] fix SanaMultiscaleLinearAttention apply_quadratic_atten…

…tion bf16 (#10595)

* autoencoder_dc tiling

* add tiling and slicing support in SANA pipelines

* create variables for padding length because the line becomes too long

* add tiling and slicing support in pag SANA pipelines

* revert changes to tile size

* make style

* add vae tiling test

* fix SanaMultiscaleLinearAttention apply_quadratic_attention bf16

---------

Co-authored-by: Aryan <[email protected]>

src/diffusers/models/attention_processor.py

@@ -899,7 +899,7 @@ def apply_quadratic_attention(self, query: torch.Tensor, key: torch.Tensor, valu
         scores = torch.matmul(key.transpose(-1, -2), query)
         scores = scores.to(dtype=torch.float32)
         scores = scores / (torch.sum(scores, dim=2, keepdim=True) + self.eps)
-        hidden_states = torch.matmul(value, scores)
+        hidden_states = torch.matmul(value, scores.to(value.dtype))
         return hidden_states
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: