diff --git a/tests/lora/test_lora_weights.py b/tests/lora/test_lora_weights.py
new file mode 100644
index 0000000000..ff13214793
--- /dev/null
+++ b/tests/lora/test_lora_weights.py
@@ -0,0 +1,52 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import pytest
+import torch
+
+from vllm.lora.lora_weights import LoRALayerWeights, PackedLoRALayerWeights
+
+RANK, ALPHA = 16, 32
+# PEFTHelper stores alpha/sqrt(r) for use_rslora adapters.
+RSLORA_SCALING = ALPHA / RANK**0.5
+
+
+@pytest.fixture(autouse=True)
+def should_do_global_cleanup_after_test():
+    # Pure-python weight packing; skip GPU/ray cleanup.
+    return False
+
+
+def moe_loras(scaling: float, stacked: bool = False) -> list[LoRALayerWeights]:
+    a_shape = (3, RANK, 4) if stacked else (RANK, 4)
+    b_shape = (3, 6, RANK) if stacked else (6, RANK)
+    return [
+        LoRALayerWeights(
+            module_name=f"experts.w{i}",
+            rank=RANK,
+            lora_alpha=ALPHA,
+            lora_a=torch.randn(*a_shape),
+            lora_b=torch.randn(*b_shape),
+            scaling=scaling,
+        )
+        for i in (1, 2, 3)
+    ]
+
+
+@pytest.mark.parametrize(
+    "packer,stacked",
+    [
+        (PackedLoRALayerWeights.pack_moe, False),
+        (PackedLoRALayerWeights.pack_moe_stacked, True),
+    ],
+)
+def test_moe_packing_uses_stored_scaling(packer, stacked: bool):
+    packed = packer(moe_loras(RSLORA_SCALING, stacked), "experts")
+    assert packed.scaling == pytest.approx([RSLORA_SCALING] * 3)
+
+
+def test_non_gated_moe_keeps_w3_unscaled():
+    packed = PackedLoRALayerWeights.pack_moe(
+        moe_loras(RSLORA_SCALING), "experts", is_non_gated_moe=True
+    )
+    assert packed.scaling[:2] == pytest.approx([RSLORA_SCALING] * 2)
+    assert packed.scaling[2] == 1.0
diff --git a/vllm/lora/lora_weights.py b/vllm/lora/lora_weights.py
index 38a5875525..bb241db77e 100644
--- a/vllm/lora/lora_weights.py
+++ b/vllm/lora/lora_weights.py
@@ -201,8 +201,9 @@ class PackedLoRALayerWeights(LoRALayerWeights):
         w1_lora_b = torch.stack(w1_lora_b_lst, dim=0)  # (num_experts,output_size,rank)
         w2_lora_b = torch.stack(w2_lora_b_lst, dim=0)
 
-        # All w1, w2, w3 have the same scaling factor.
-        scaling = lora_alpha / rank
+        # All w1, w2, w3 have the same scaling factor. Use the per-adapter
+        # scaling (e.g. alpha/sqrt(r) for rsLoRA) instead of alpha/rank.
+        scaling = first_lora.scaling
         last_scaling = scaling
 
         if is_non_gated_moe:
@@ -250,7 +251,7 @@ class PackedLoRALayerWeights(LoRALayerWeights):
         assert w1_lora is not None and w2_lora is not None and w3_lora is not None
         rank = w1_lora.rank
         lora_alpha = w1_lora.lora_alpha
-        scaling = lora_alpha / rank
+        scaling = w1_lora.scaling
         return cls(
             module_name,
             rank,