fix(model): rewrite seq_cls architectures to *ForSequenceClassification

[RerankerGuo]

Summary

Fixes #9704.

When fine-tuning a VLM or LM with task_type=seq_cls (or reranker), the seq_cls patcher in _patch_sequence_classification monkey-patches a score head onto the generation model class without swapping the class itself. transformers.PreTrainedModel.save_pretrained then writes model.__class__.__name__ into config.json['architectures'], so the on-disk checkpoint advertises e.g. Qwen3VLForConditionalGeneration while shipping a score head, num_labels, id2label, and problem_type=multi_label_classification.

Downstream vLLM deployment reads architectures to pick the model class, finds the generation class, and rejects the checkpoint — even though PtEngine inference works correctly. The reporter in #9704 confirmed all four of score.weight: [20, 4096], id2label, label2id, and problem_type are written correctly; only architectures is wrong.

Fix

Rewrite model.config.architectures in-place to the matching *ForSequenceClassification class at the same place the score head is attached, so every save path (trainer _save_model, save_checkpoint, export-quant, peft merge-and-unload) writes the right value automatically. The existing vllm_engine.py:347-353 arch_mapping override is left in place — it remains useful for users who already have bad checkpoints on disk.

The new helper _seq_cls_architectures handles the suffix rewrite idempotently and leaves unknown / custom architectures untouched (so we never advertise a class that does not exist in transformers).

Mapping coverage

Source class	Rewritten to
*ForConditionalGeneration (Qwen3-VL, Qwen2-VL, InternVL, LLaVA, GLM4V, …)	*ForSequenceClassification
*ForCausalLM (Qwen2, Llama, Mistral, Mixtral, Yi, …)	*ForSequenceClassification
Already *ForSequenceClassification	unchanged (idempotent)
Custom / unknown suffix	unchanged
[] / None	unchanged

Tests

Added TestSeqClsArchitecturesRewrite in tests/general/test_model.py covering:

• generation → seq_cls rewrite for both ForCausalLM and ForConditionalGeneration suffixes
• idempotence on already-seq_cls classes
• multi-arch lists
• empty / None inputs
• unknown / custom architectures left untouched

Local sanity checks (10/10 PASS) and syntax checks on both files PASS. Full ms-swift unit / smoke CI is needed to validate end-to-end on real seq_cls training runs; that will run on the upstream CI after this PR is opened.

Out of scope

• Megatron seq_cls architectures rewrite — Megatron has its own seq_cls head code path (swift/megatron/utils/convert_utils.py:283, swift/megatron/model/utils.py:47); happy to follow up with a separate PR if maintainers want it.
• vllm_engine.py:347-353 arch_mapping cleanup — kept for backward-compat with existing bad checkpoints.

🤖 Generated with Claude Code