fix: guard model.half() with dtype check in all rerankers

[nedeadinside]

Calling .half() on an already-FP16 model raises an error, causing rerankers with use_fp16=True to crash on every request after the first. Add a dtype guard so the conversion only runs when needed:

if self.use_fp16 and next(self.model.parameters()).dtype != torch.float16:
self.model.half()

Fixes BaseReranker, BaseLLMReranker, LightweightLLMReranker, LayerWiseLLMReranker, and MatroyshkaReranker.

[Copilot]

Pull request overview

This PR prevents reranker inference from repeatedly calling model.half() when the loaded model is already FP16, which can crash subsequent requests when use_fp16=True. It adds a dtype guard in the compute_score_single_gpu path across the core rerankers (and the Matroyshka research reranker).

Changes:

• Guard self.model.half() behind a dtype check (!= torch.float16) to avoid redundant FP16 conversion.
• Apply the guard consistently across encoder-only, decoder-only, and Matroyshka reranker implementations.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
research/Matroyshka_reranker/inference/rank_model.py	Add FP16 dtype guard before calling model.half() in Matroyshka reranker inference.
FlagEmbedding/inference/reranker/encoder_only/base.py	Add FP16 dtype guard before calling model.half() in the encoder-only reranker.
FlagEmbedding/inference/reranker/decoder_only/lightweight.py	Add FP16 dtype guard before calling model.half() in the lightweight LLM reranker.
FlagEmbedding/inference/reranker/decoder_only/layerwise.py	Add FP16 dtype guard before calling model.half() in the layer-wise LLM reranker.
FlagEmbedding/inference/reranker/decoder_only/base.py	Add FP16 dtype guard before calling model.half() in the base decoder-only LLM reranker.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.