[quantization] Properly ignore quantization for layers excluded in quant_config (#11205)

python/sglang/srt/layers/moe/fused_moe_triton/layer.py

@@ -207,15 +207,11 @@ class FusedMoE(torch.nn.Module):
             gemm1_clamp_limit=gemm1_clamp_limit,
         )
 
-        if quant_config is None:
-            self.quant_method: FusedMoEMethodBase = UnquantizedFusedMoEMethod(
-                self.use_triton_kernels
-            )
-        else:
-            self.quant_method: FusedMoEMethodBase = quant_config.get_quant_method(
-                self, prefix
-            )
-        assert self.quant_method is not None
+        self.quant_method: Optional[FusedMoEMethodBase] = None
+        if quant_config is not None:
+            self.quant_method = quant_config.get_quant_method(self, prefix)
+        if self.quant_method is None:
+            self.quant_method = UnquantizedFusedMoEMethod(self.use_triton_kernels)
 
         self.quant_method.create_weights(
             layer=self,

python/sglang/srt/layers/quantization/quark/quark.py

@@ -65,7 +65,9 @@ class QuarkConfig(QuantizationConfig):
         if should_ignore_layer(
             prefix, ignore=exclude_layers, fused_mapping=self.packed_modules_mapping
         ):
-            return UnquantizedLinearMethod()
+            if isinstance(layer, LinearBase):
+                return UnquantizedLinearMethod()
+            return None
 
         if isinstance(layer, LinearBase):
             scheme = self.get_scheme(layer=layer, layer_name=prefix)