Fix missing .unsqueeze(1) in attention mask for HunyuanVideo VAE

Minisal · Minisal · commit d1b69c90593a · 2026-02-13T11:28:37.000+08:00
diff --git a/src/diffusers/models/autoencoders/autoencoder_kl_hunyuanvideo15.py b/src/diffusers/models/autoencoders/autoencoder_kl_hunyuanvideo15.py
@@ -145,7 +145,7 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
             frames, height * width, query.dtype, query.device, batch_size=batch_size
         )
 
-        x = nn.functional.scaled_dot_product_attention(query, key, value, attn_mask=attention_mask)
+        x = nn.functional.scaled_dot_product_attention(query, key, value, attn_mask=attention_mask.unsqueeze(1))
 
         # batch_size, 1, frames * height * width, channels
 

Original file line number	Diff line number	Diff line change
`@@ -145,7 +145,7 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:`
`145`	`145`	`frames, height * width, query.dtype, query.device, batch_size=batch_size`
`146`	`146`	`)`
`147`	`147`
`148`		`- x = nn.functional.scaled_dot_product_attention(query, key, value, attn_mask=attention_mask)`
	`148`	`+ x = nn.functional.scaled_dot_product_attention(query, key, value, attn_mask=attention_mask.unsqueeze(1))`
`149`	`149`
`150`	`150`	`# batch_size, 1, frames * height * width, channels`
`151`	`151`