From 8481327d124f6fda63579e3dba8742d98dbeae5c Mon Sep 17 00:00:00 2001 From: electroglyph Date: Tue, 16 Dec 2025 02:32:58 -0800 Subject: [PATCH] force disable FP8 loading --- unsloth/models/sentence_transformer.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/unsloth/models/sentence_transformer.py b/unsloth/models/sentence_transformer.py index 91b18c46a7..36bc14e203 100644 --- a/unsloth/models/sentence_transformer.py +++ b/unsloth/models/sentence_transformer.py @@ -44,7 +44,6 @@ class FastSentenceTransformer(FastModel): max_lora_rank = 64, disable_log_stats = True, qat_scheme = None, - load_in_fp8 = False, unsloth_tiled_mlp = False, pooling_mode = "mean", **kwargs, @@ -65,6 +64,10 @@ class FastSentenceTransformer(FastModel): if "add_pooling_layer" not in kwargs: kwargs["add_pooling_layer"] = False + # forces fp8 to be False since it's not supported + kwargs.pop("load_in_fp8", None) + load_in_fp8 = False + # this is a fix for Snowflake/snowflake-arctic-embed-l-v2.0 # it has pooler weights which we don't care about for training, # however unsloth throws an exception if "UNSLOTH_WARN_UNINITIALIZED" == 1 and it sees unused weights