fix: load embed_tokens and norm for ColQwen2/ColQwen2_5 on transformers 5.x

[whybe-choi]

Summary

Fixes #413.

On transformers 5.x, Qwen2VLModel.__init__ nests the text backbone under language_model, so the expected keys are language_model.embed_tokens, language_model.layers.*, language_model.norm. Released Qwen2-VL checkpoints store these as model.embed_tokens, model.layers.*, model.norm.

ColQwen2._checkpoint_conversion_mapping only remapped model.layers, so embed_tokens and norm were silently dropped and randomly re-initialized when loading from a Qwen2-VL checkpoint. ColQwen2_5 had the same incomplete mapping and was affected identically.

Changes

Add the missing rules to both modeling_colqwen2.py and modeling_colqwen2_5.py:

_checkpoint_conversion_mapping = {
    r"^base_model\.model\.custom_text_proj": "custom_text_proj",
    r"^model\.layers": "language_model.layers",
    r"^model\.embed_tokens": "language_model.embed_tokens",
    r"^model\.norm": "language_model.norm",
}

This is backward-compatible: the rules only match keys starting with model., so checkpoints already saved in the language_model.* layout are unaffected.

Result

Before, language_model.embed_tokens.weight and language_model.norm.weight were reported MISSING while their source tensors were UNEXPECTED. After the fix only the expected custom_text_proj.{weight,bias} (the ColBERT projection head trained from scratch) remain MISSING.

[QuentinJGMace]

Thanks for finding this and the fix ! Could you also update the changelog ?

Otherwise it looks good to me.

[whybe-choi]

Done — added the changelog entry. Thanks!