From da42e16505a0d54d87b77be295592f315a833d05 Mon Sep 17 00:00:00 2001 From: electroglyph Date: Tue, 16 Dec 2025 04:04:39 -0800 Subject: [PATCH] fix _save_pretrained_merged --- unsloth/models/sentence_transformer.py | 11 +++++++++++ 1 file changed, 11 insertions(+) diff --git a/unsloth/models/sentence_transformer.py b/unsloth/models/sentence_transformer.py index 0be988483e..45830de733 100644 --- a/unsloth/models/sentence_transformer.py +++ b/unsloth/models/sentence_transformer.py @@ -232,6 +232,17 @@ class FastSentenceTransformer(FastModel): except: pass + # remove quantization_config to prevent saving 4bit config for a 16bit merged model + config = self[0].auto_model.config + original_quantization_config = getattr(config, "quantization_config", None) + if original_quantization_config is not None: + del config.quantization_config + + config.save_pretrained(save_directory) + + if original_quantization_config is not None: + config.quantization_config = original_quantization_config + # save merged weights tokenizer = kwargs.pop("tokenizer", self.tokenizer) self[0].auto_model.save_pretrained_merged(save_directory, tokenizer=tokenizer, **kwargs)