- Add proper tensor reshaping in vae_sage_attention() wrapper
- Convert VAE format (B, C, H, W) to SageAttention format (B, seq_len, dim)
- Add fallback mechanism to pytorch attention if SageAttention fails
- Enable SageAttention selection in vae_attention() function
- Fixes ValueError: too many values to unpack (expected 3) error
- Maintains compatibility with existing VAE attention functions
This resolves the tensor shape incompatibility between VAE spatial attention
and SageAttention sequence attention, allowing VAE operations to benefit from
SageAttention optimization when --use-sage-attention flag is enabled.
0b0753e708