MiniMax H3 transformer does not expose attention masks for padded joint sequences #14870

Problem

MiniMaxH3TransformerBlock.forward accepts an attention_mask, but MiniMaxH3Transformer3DModel.forward does not. Standard Ulysses context parallelism needs masked padding when the packed text/video/audio sequence length is not divisible by the parallel degree. The copied forward used by verl-omni #635 is a temporary workaround that may break on a Diffusers upgrade.

Reproduction

On Diffusers 0.40.0 or main at 4295ee3:

import inspect
from diffusers import MiniMaxH3Transformer3DModel

assert "attention_mask" in inspect.signature(MiniMaxH3Transformer3DModel.forward).parameters

The assertion fails because the argument is absent.

Expected behavior

Accept an optional boolean key-padding mask and pass it to every transformer block, including during gradient checkpointing. Calls without a mask should remain unchanged.

Related work

Fix: Diffusers #14868. Downstream workaround: verl-omni #635. This is separate from the Ulysses-Anything backward work in #14834.