attention dispatcher assumes wrong attributes for flash attn kernel from hub

[bghira]

Describe the bug

The breakage was in the active diffusers checkout, not in kernels. kernels-community/flash-attn2 exposes:

• flash_attn_varlen_func
• flash_attn_varlen_qkvpacked_func
• flash_attn_interface._flash_attn_varlen_forward
• flash_attn_interface._flash_attn_varlen_backward

But diffusers was trying to resolve:

• flash_attn_interface._wrapped_flash_attn_varlen_forward
• flash_attn_interface._wrapped_flash_attn_varlen_backward

Those wrapped* attrs do not exist in the current kernels package. I patched src/diffusers/models/attention_dispatch.py to use the actual _flash_attn_varlen_forward/...backward names, which allowed it to proceed.

Reproduction

No separate MRE

Logs

System Info

Not needed

Who can help?

No response