Ask transformers whether a model supports context parallelism

[qgallouedec]

Follow-up to #4177, and depends on huggingface/transformers#48442. Draft until that merges.

#4177 pattern-matches layer_types and sliding_window from inside accelerate. @SunMarc noted in review that this belongs in transformers, and huggingface/transformers#48442 adds supports_context_parallel for it. This swaps the inline check for that property.

It also picks up something the config check could not express: a model can set _supports_context_parallel = False to rule context parallelism out whatever its config, which is the case for gpt-oss (attention sinks need a custom kernel that SDPA cannot provide).

Defaults to allowed when the property is absent, so an older transformers behaves exactly as it does today.

Leaves _refuse_recurrent_layers_under_sequence_parallelism alone on purpose. supports_context_parallel is also False for sliding-window models, and I have not verified that those are broken under Ulysses, so reusing it there would refuse models that may well work.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[SunMarc]

That would be the better solution but let's see if this is worth adding it in transformers itself.

[qgallouedec]

Holding this until #4209 lands, then rebuilding it on top, two changes:

• The getattr(model, "supports_context_parallel", True) default admits sliding-window / chunked models on every transformers without the property, i.e. every release through 5.16.1, undoing Refuse context parallelism for models with sliding-window or chunked attention layers #4177. use ..._TRANSFORMERS_VERSION = "5.17.0" constant and compare_versions("transformers", ">=", ...), the property on new transformers, the current inspection otherwise.
• Add an FSDP2-native Ulysses sequence parallelism backend #4209 factors that inspection into _non_full_attention_layer_types and reuses it for Ulysses, where the CP property doesn't apply (SP allows sliding-window with flash attention). So this PR only swaps the CP side, and the helper stays.