From 61da34ad0b8a0f9b844ea123b1f206fad419c793 Mon Sep 17 00:00:00 2001
From: Changyi Yang <112288487+ChangyiYang@users.noreply.github.com>
Date: Mon, 16 Feb 2026 02:39:45 -0500
Subject: [PATCH] [diffusion] fix: fix LoRA weight snapshot aliasing in unmerge
 logic (#18883)

---
 python/sglang/multimodal_gen/runtime/layers/lora/linear.py | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/python/sglang/multimodal_gen/runtime/layers/lora/linear.py b/python/sglang/multimodal_gen/runtime/layers/lora/linear.py
index a062479d1..e83f75a49 100644
--- a/python/sglang/multimodal_gen/runtime/layers/lora/linear.py
+++ b/python/sglang/multimodal_gen/runtime/layers/lora/linear.py
@@ -48,7 +48,9 @@ class BaseLayerWithLoRA(nn.Module):
         self.base_layer: nn.Module = base_layer
 
         self.merged: bool = False
-        self.cpu_weight = base_layer.weight.to("cpu")
+        # Immutable base-weight snapshot; `to("cpu")` may alias CPU storage.
+        # Use `clone()` so merge updates cannot mutate this backup tensor.
+        self.cpu_weight = base_layer.weight.detach().to("cpu").clone()
         # indicates adapter weights don't contain this layer
         # (which shouldn't normally happen, but we want to separate it from the case of erroneous merging)
         # Default to True to prevent using uninitialized weights; set to False when weights are loaded