diff --git a/aphrodite/lora/layers/utils.py b/aphrodite/lora/layers/utils.py
index 9873f2e029..3a64b910b3 100644
--- a/aphrodite/lora/layers/utils.py
+++ b/aphrodite/lora/layers/utils.py
@@ -69,6 +69,10 @@ def _get_lora_device(base_layer: nn.Module) -> torch.device:
     # MoE GPTQ/AWQ/GGUF
     elif hasattr(base_layer, "w2_qweight"):
         return base_layer.w2_qweight.device
+    # EXL3 (exllamav3) — CUDA-only; tensors are placed on the current device
+    # in Exl3LinearMethod.process_weights_after_loading.
+    elif hasattr(base_layer, "trellis"):
+        return torch.device("cuda", torch.accelerator.current_device_index())
     else:
         raise ValueError(f"Unsupported base layer: {base_layer}")