jinaai

/

xlm-roberta-flash-implementation

Like 34

Follow

Jina AI 2.19k

Fixes #62. Heads up: I prepared this with AI assistance (Claude Code) and reviewed and tested it myself.

XLMRobertaLoRA.__init__ doesn't call self.post_init(), which transformers v5 requires (#60 added it to the classes in modeling_xlm_roberta.py, but not here). So loading jinaai/jina-embeddings-v3 on transformers>=5 fails with:

AttributeError: 'XLMRobertaLoRA' object has no attribute 'all_tied_weights_keys'

This adds self.post_init() at the end of XLMRobertaLoRA.__init__.

Testing (RTX PRO 6000; SentenceTransformer("jinaai/jina-embeddings-v3", trust_remote_code=True).encode(..., task="text-matching") on 4 sentences)

• The load_trained_adapters=False path (fresh adapters on a pretrained XLMRobertaModel) gives a bit-identical state dict to the current code on 4.57.6 (seeded, 492/492 tensors equal), so post_init() doesn't re-initialize anything there.
• vLLM's test_matryoshka HF-vs-vLLM comparison for jina-embeddings-v3 passes with this change on transformers 5.17.0 (it's currently skipped in vLLM because of this error).