From c35eb0fdee490f9844dad08fc3fe97cb4a124ee6 Mon Sep 17 00:00:00 2001 From: Lianmin Zheng Date: Mon, 5 Jan 2026 01:34:30 -0800 Subject: [PATCH] [Auto Sync] Update loader.py (20260105) (#16478) Co-authored-by: github-actions[bot] Co-authored-by: xiuyuli-xai --- python/sglang/srt/model_loader/loader.py | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/python/sglang/srt/model_loader/loader.py b/python/sglang/srt/model_loader/loader.py index 3f624aca2..deaa1fff2 100644 --- a/python/sglang/srt/model_loader/loader.py +++ b/python/sglang/srt/model_loader/loader.py @@ -1240,6 +1240,12 @@ class DummyModelLoader(BaseModelLoader): for _, module in model.named_modules(): quant_method = getattr(module, "quant_method", None) if quant_method is not None: + # Skip FusedMoE layers already quantized during init (FP8 or FP4) + if ( + hasattr(module, "is_weights_quantized") + and module.is_weights_quantized() + ): + continue quant_method.process_weights_after_loading(module) # NOTE(woosuk): For accurate performance evaluation, we assign