diff --git a/vllm/compilation/passes/fusion/allreduce_rms_fusion.py b/vllm/compilation/passes/fusion/allreduce_rms_fusion.py
index 6cb0c8f49f3d..83423c02a6f1 100644
--- a/vllm/compilation/passes/fusion/allreduce_rms_fusion.py
+++ b/vllm/compilation/passes/fusion/allreduce_rms_fusion.py
@@ -939,7 +939,7 @@ def replacement(self):
         def _replacement(
             input: torch.Tensor, weight: torch.Tensor
         ) -> tuple[torch.Tensor, torch.Tensor]:
-            residual = torch.empty_like(input)
+            residual = torch.zeros_like(input)
             allreduce = self.FUSED_AR_RMSNORM_OP(
                 input_=input,
                 residual=residual,
diff --git a/vllm/model_executor/layers/layernorm.py b/vllm/model_executor/layers/layernorm.py
index a5d4e4db79fe..f9897a22c8d5 100644
--- a/vllm/model_executor/layers/layernorm.py
+++ b/vllm/model_executor/layers/layernorm.py
@@ -90,7 +90,11 @@ def forward_native(
                 x,
                 self.weight.data if self.pass_weight else None,
                 self.variance_epsilon,
-                self.variance_size_override,
+                *(
+                    (self.variance_size_override,)
+                    if self.variance_size_override is not None
+                    else ()
+                ),
             )
         else:
             return ir.ops.fused_add_rms_norm.maybe_inplace(
@@ -98,7 +102,11 @@ def forward_native(
                 residual,
                 self.weight.data if self.pass_weight_add else None,
                 self.variance_epsilon,
-                self.variance_size_override,
+                *(
+                    (self.variance_size_override,)
+                    if self.variance_size_override is not None
+                    else ()
+                ),
             )
 
     def forward_cuda(