From 7a607c49009de956b6dd8c0e3055825ae8efb4a6 Mon Sep 17 00:00:00 2001 From: tjp_zju Date: Mon, 16 Feb 2026 00:31:42 +0800 Subject: [PATCH] fix_get_quant_method_in_fused_moe_condition (#18459) Signed-off-by: tom-zju Co-authored-by: Peng Zhang --- python/sglang/srt/layers/quantization/moe_wna16.py | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/python/sglang/srt/layers/quantization/moe_wna16.py b/python/sglang/srt/layers/quantization/moe_wna16.py index 531e4271f..ce6397dd1 100644 --- a/python/sglang/srt/layers/quantization/moe_wna16.py +++ b/python/sglang/srt/layers/quantization/moe_wna16.py @@ -18,7 +18,10 @@ from sglang.srt.layers.quantization.base_config import ( QuantizeMethodBase, ) from sglang.srt.layers.quantization.gptq import GPTQConfig, GPTQMarlinConfig -from sglang.srt.layers.quantization.unquant import UnquantizedLinearMethod +from sglang.srt.layers.quantization.unquant import ( + UnquantizedFusedMoEMethod, + UnquantizedLinearMethod, +) from sglang.srt.utils import get_device_capability, set_weight_attrs logger = logging.getLogger(__name__) @@ -194,6 +197,8 @@ class MoeWNA16Config(QuantizationConfig): from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE if is_layer_skipped_quant(prefix, self.modules_to_not_convert): + if isinstance(layer, FusedMoE): + return UnquantizedFusedMoEMethod() return UnquantizedLinearMethod() elif isinstance(layer, LinearBase):