diff --git a/vllm/model_executor/layers/fused_moe/oracle/mxfp4.py b/vllm/model_executor/layers/fused_moe/oracle/mxfp4.py
index cd4b30b772..e6e9d17925 100644
--- a/vllm/model_executor/layers/fused_moe/oracle/mxfp4.py
+++ b/vllm/model_executor/layers/fused_moe/oracle/mxfp4.py
@@ -1146,8 +1146,13 @@ def convert_gpt_oss_weight_to_mxfp4_moe_kernel_format(
             weight_scale=w2_scale, flex_ctx=FlexCtx(rhs_data=w2_flex)
         )
 
+        # The original mxfp4 block scales have been swizzled into the
+        # precision configs above and are no longer read by the kernel, so
+        # drop the now-dead weight/scale Parameters to free their memory.
         del layer.w13_weight
         del layer.w2_weight
+        del layer.w13_weight_scale
+        del layer.w2_weight_scale
 
         return (
             w13_weight,
@@ -1508,8 +1513,13 @@ def convert_weight_to_mxfp4_moe_kernel_format(
             weight_scale=w2_scale, flex_ctx=FlexCtx(rhs_data=w2_flex)
         )
 
+        # The original mxfp4 block scales have been swizzled into the
+        # precision configs above and are no longer read by the kernel, so
+        # drop the now-dead weight/scale Parameters to free their memory.
         del layer.w13_weight
         del layer.w2_weight
+        del layer.w13_weight_scale
+        del layer.w2_weight_scale
 
         return (
             w13_weight,
diff --git a/vllm/model_executor/layers/quantization/mxfp4.py b/vllm/model_executor/layers/quantization/mxfp4.py
index 1b2a8a74bd..5ef5fd40d5 100644
--- a/vllm/model_executor/layers/quantization/mxfp4.py
+++ b/vllm/model_executor/layers/quantization/mxfp4.py
@@ -393,16 +393,19 @@ class GptOssMxfp4MoEMethod(FusedMoEMethodBase):
     def get_fused_moe_quant_config(
         self, layer: RoutedExperts
     ) -> FusedMoEQuantConfig | None:
-        w1_scale = layer.w13_weight_scale
-        w2_scale = layer.w2_weight_scale
         w1_bias = getattr(layer, "w13_bias", None)
         w2_bias = getattr(layer, "w2_bias", None)
 
         if self.mxfp4_backend in TRITON_BACKENDS:
+            # TRITON backends free w13/w2_weight_scale after swizzling; the
+            # swizzled scales live inside the precision configs instead.
             assert self.w13_precision_config is not None
             assert self.w2_precision_config is not None
             w1_scale = self.w13_precision_config
             w2_scale = self.w2_precision_config
+        else:
+            w1_scale = layer.w13_weight_scale
+            w2_scale = layer.w2_weight_scale
 
         return make_mxfp4_moe_quant_config(
             mxfp4_backend=self.mxfp4_backend,
@@ -738,17 +741,20 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
         self,
         layer: RoutedExperts,
     ) -> FusedMoEQuantConfig | None:
-        w1_scale = layer.w13_weight_scale
-        w2_scale = layer.w2_weight_scale
         w1_bias = getattr(layer, "w13_bias", None)
         w2_bias = getattr(layer, "w2_bias", None)
         swiglu_limit = getattr(layer, "swiglu_limit", None)
 
         if self.mxfp4_backend in TRITON_BACKENDS:
+            # TRITON backends free w13/w2_weight_scale after swizzling; the
+            # swizzled scales live inside the precision configs instead.
             assert self.w13_precision_config is not None
             assert self.w2_precision_config is not None
             w1_scale = self.w13_precision_config
             w2_scale = self.w2_precision_config
+        else:
+            w1_scale = layer.w13_weight_scale
+            w2_scale = layer.w2_weight_scale
 
         return make_mxfp4_moe_quant_config(
             mxfp4_backend=self.mxfp4_backend,