[Feature] Reuse Ulysses shard lengths across MiniMax-H3 blocks #15132

MiniMax-H3 keeps the same packed text/video/audio sequence throughout its transformer blocks. With Ulysses Anything, each attention call still gathers shard lengths separately for Q, K, V and the output exchange. This repeats metadata communication even though the layout is unchanged within that forward.

On main at aaf57d03, the Q/K/V helper calls gather_size_by_comm, as does the output helper. H3 then repeats attention across its block stack.

Could H3 share an immutable shard-length plan within one inference forward? The plan should be rebuilt on every call, preserve collective ordering on all ranks, and leave the default and token-refiner paths unchanged. It should not become a cache shared between requests.

The proposed opt-in in #15064 is:

with torch.inference_mode():
    output = model(**inputs, attention_kwargs={"reuse_ulysses_metadata": True})

The existing two-process CPU/Gloo checks for that implementation, recorded on base d42ebf9c, matched the default tiny-H3 outputs for packed lengths 18 → 20 → 18 while using fewer all-gathers. Uneven attention checks covered 9 → 10 → 9; inconsistent full layouts raised on both ranks, and the original parallel configuration stayed unchanged.

DIFFUSERS_TEST_DEVICE=cpu PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 OMP_NUM_THREADS=1 \
  python -m pytest tests/models/transformers/test_models_transformer_minimax_h3.py::TestMiniMaxH3SequencePlan -q

These are prior CPU/Gloo results for the proposed implementation, not a new run on current main. CUDA/NCCL, compile, isolated latency, full-checkpoint generation and output quality remain unvalidated; no speedup is claimed.

#14871/#14834 cover Ulysses Anything backward and metadata within an attention call. This request concerns reuse across H3 blocks. Would you prefer this scope in #15064 or an extension of that lower-level work? Maintainer coordination is pending. AI assistance was used to prepare the implementation and this proposal.