From 9a48eef89a777fc7039d9765fb2f6fc618216948 Mon Sep 17 00:00:00 2001
From: Alejandro Paredes La Torre
 <99832715+AlejandroParedesLT@users.noreply.github.com>
Date: Sat, 11 Jul 2026 17:13:50 -0700
Subject: [PATCH] [Bugfix][LoRA] Support ark_linear base layer in
 _get_lora_device (#47690)

Signed-off-by: AlejandroParedesLT <alejandroparedeslatorre@gmail.com>
---
 tests/lora/test_layers_utils.py | 39 +++++++++++++++++++++++++++++++++
 vllm/lora/layers/utils.py       |  3 +++
 2 files changed, 42 insertions(+)
 create mode 100644 tests/lora/test_layers_utils.py

diff --git a/tests/lora/test_layers_utils.py b/tests/lora/test_layers_utils.py
new file mode 100644
index 0000000000..0087519a33
--- /dev/null
+++ b/tests/lora/test_layers_utils.py
@@ -0,0 +1,39 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import pytest
+import torch
+from torch import nn
+
+from vllm.lora.layers.utils import _get_lora_device
+
+pytestmark = pytest.mark.skip_global_cleanup
+
+
+def _param() -> nn.Parameter:
+    return nn.Parameter(torch.empty(1), requires_grad=False)
+
+
+def test_get_lora_device_unquantized():
+    base_layer = nn.Module()
+    base_layer.weight = _param()
+    assert _get_lora_device(base_layer) == base_layer.weight.device
+
+
+def test_get_lora_device_gptq_awq():
+    base_layer = nn.Module()
+    base_layer.qweight = _param()
+    assert _get_lora_device(base_layer) == base_layer.qweight.device
+
+
+def test_get_lora_device_ark_linear():
+    base_layer = nn.Module()
+    base_layer.ark_linear = nn.Module()
+    base_layer.ark_linear.qweight = _param()
+    assert _get_lora_device(base_layer) == base_layer.ark_linear.qweight.device
+
+
+def test_get_lora_device_unsupported_raises():
+    base_layer = nn.Module()
+    with pytest.raises(ValueError, match="Unsupported base layer"):
+        _get_lora_device(base_layer)
diff --git a/vllm/lora/layers/utils.py b/vllm/lora/layers/utils.py
index 17c21d36f5..b4cde2564b 100644
--- a/vllm/lora/layers/utils.py
+++ b/vllm/lora/layers/utils.py
@@ -57,6 +57,9 @@ def _get_lora_device(base_layer: nn.Module) -> torch.device:
     # GPTQ/AWQ
     elif hasattr(base_layer, "qweight"):
         return base_layer.qweight.device
+    # INC WNA16 (AutoRound)
+    elif hasattr(base_layer, "ark_linear"):
+        return base_layer.ark_linear.qweight.device
     # MoE layer
     elif hasattr(base_layer, "w2_weight"):
         return base_layer.w2_weight.device