From e73632f841c77d90400f8840dc4e68ef92f40a88 Mon Sep 17 00:00:00 2001 From: electroglyph Date: Sat, 10 Jan 2026 00:00:11 -0800 Subject: [PATCH] do QAT preparation for fast path --- unsloth/models/sentence_transformer.py | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/unsloth/models/sentence_transformer.py b/unsloth/models/sentence_transformer.py index 144dae9bd8..bede31a28e 100644 --- a/unsloth/models/sentence_transformer.py +++ b/unsloth/models/sentence_transformer.py @@ -1399,6 +1399,12 @@ class FastSentenceTransformer(FastModel): # Apply PEFT directly (not through FastModel) peft_model = peft_get_peft_model(inner_model, lora_config) + # Apply QAT if specified + qat_scheme = kwargs.get("qat_scheme", None) + if qat_scheme is not None: + from ._utils import _prepare_model_for_qat + peft_model = _prepare_model_for_qat(peft_model, qat_scheme) + # Determine compile mode (only if not using gradient checkpointing) compile_mode = getattr(model, "_compile_mode", "default") # Re-enable torch.compile if gradient checkpointing was requested but couldn't be enabled