Write the aligned token ids to the text config - #7315

This PR writes the aligned special token ids to the text config instead of the top-level config.

Part of #7093. For a composite model the id the trainer wrote was not the one Trainer reads, so align_special_tokens realigned it at train time and logged it; writing through get_text_config() leaves nothing to realign.

Motivation

The trainers set the tokenizer's pad and eos ids on the model configs before training, so that the alignment Trainer performs at train time finds nothing to change. huggingface/transformers#48847 moved that alignment to read and write model.config.get_text_config(), because a composite model keeps its special token ids on the text sub-config and the top-level config does not forward attribute lookups.

Our writes still land on the top-level config, so for a composite model they are no longer what gets compared, and the realignment message comes back. Measured on tiny-Qwen2_5_VLForConditionalGeneration against transformers main:

TRL writes to model.config                       -> Updated tokens: {'pad_token_id': 151643}
TRL writes to model.config.get_text_config()     -> no message

Solution

Write to model.config.get_text_config() at each site that mirrors a token id for that purpose. get_text_config() returns self for a non-composite config, so this is a no-op for text models and needs no version guard: the method predates our transformers>=4.56.2 floor.

The generation config is untouched, since it is not composite.

Changes

• Write the mirrored pad token id to the text config in the SFT, DPO, KTO, GRPO, RLOO, Distillation and Reward trainers
• Write the mirrored eos token id to the text config in the SFT and Reward trainers

CC: @qgallouedec and @lewtun for visibility.