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

[Minisal]

What does this PR do?

This PR fixes a dimension mismatch in the attention mask handling for the HunyuanVideo1.5 VAE implementation in autoencoder_kl_hunyuanvideo15.py.

Problem

In the current implementation (line 148 of autoencoder_kl_hunyuanvideo15.py), the attention_mask is passed directly to scaled_dot_product_attention without reshaping:

x = nn.functional.scaled_dot_product_attention(query, key, value, attn_mask=attention_mask)

However, the original HuanyuanVideo implementation (line 211 of huanyuanvideo_15_vae.py) correctly applies .unsqueeze(1) to the mask:

h_ = nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=attention_mask.unsqueeze(1))

This ensures the attention mask has the correct shape for batched attention computation. Without this, the mask dimensions may not align with the query/key/value tensors, leading to potential runtime errors or incorrect attention behavior.

Fix

Add .unsqueeze(1) to the attention_mask argument in the scaled_dot_product_attention call to match the original implementation.

---

Changes

• File: src/diffusers/models/autoencoder_kl_hunyuanvideo15.py
• Line: 148
• Before:

  x = nn.functional.scaled_dot_product_attention(query, key, value, attn_mask=attention_mask)

• After:

  x = nn.functional.scaled_dot_product_attention(query, key, value, attn_mask=attention_mask.unsqueeze(1))

---

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@yiyixuxu
@asomoza
@sayakpaul
@DN6

[sayakpaul]

Can you provide a reproducer?