[docs] add security notes to NPU Docker overviews (#10883)

docker/docker-npu/OVERVIEW.md

@@ -184,6 +184,15 @@ To build an image with Docker Compose without starting a container, use `docker
 - Legacy NPU tags are replaced by the `latest-<910b|a3|950>-<ubuntu|openeuler>` format.
 - Validate the exact driver, firmware, CANN, and SoC combination before production deployment.
 
+## Security Risks
+When running containers with these images, be aware of the following security risks:
+
+- Running as the `root` user: Containers run as the `root` user by default, which can introduce security risks. In production environments, it is recommended to create a non-privileged user to run the application.
+
+- Lack of CPU and memory resource limits: Not setting resource limits may cause a container to consume excessive system resources and affect host performance. It is recommended to use the `--cpus` and `--memory` parameters to limit resource usage.
+
+- Device use with `rwm` permissions: The NPU device is assigned read, write, and `mknod` permissions. Although these are required for functionality, the scope of permissions should be carefully evaluated in security-sensitive environments.
+
 ## License and Disclaimer
 
 LlamaFactory is distributed under the [Apache License 2.0](../../LICENSE).

docker/docker-npu/OVERVIEW.zh.md

@@ -184,6 +184,16 @@ docker compose --profile 950-openeuler up -d
 - 旧式 NPU tag 已由 `latest-<910b|a3|950>-<ubuntu|openeuler>` 格式取代。
 - 正式部署前，请验证具体驱动、固件、CANN 和 SoC 组合的兼容性。
 
+## 安全风险
+
+在运行容器时，需要注意以下安全风险：
+
+- 使用 `root` 用户运行：容器默认以 `root` 用户身份运行，可能带来安全隐患。建议在生产环境中创建非特权用户来运行应用程序。
+
+- 缺少 CPU 和内存资源限制：未设置资源限制可能导致容器消耗过多系统资源，影响宿主机性能。建议使用 `--cpus` 和 `--memory` 参数限制资源使用。
+
+- 设备使用 `rwm` 权限：为 NPU 设备分配了读写和 `mknod` 权限，虽然功能运行需要，但在安全敏感环境中应谨慎评估权限范围。
+
 ## 许可证与免责声明
 
 LlamaFactory 基于 [Apache License 2.0](../../LICENSE) 发布。

src/llamafactory/model/patcher.py

@@ -86,12 +86,9 @@ def _check_fla_dependencies() -> None:
 
 
 def patch_qwen3_5_forward_npu(model: "PreTrainedModel") -> None:
-    """Patch for Qwen3.5 models on NPU by importing torch_npu to enable torch.cuda compatibility.
+    """Replace the Qwen3.5 gated delta rule kernel with the NPU-compatible implementation.
 
-    On NPU, torch.cuda operations will fail unless torch_npu is imported.
-    torch_npu provides compatibility layer that maps torch.cuda calls to NPU operations.
-
-    Also replaces chunk_gated_delta_rule with NPU-compatible implementation.
+    Skip the patch unless an Ascend910-series device is detected and triton_ascend is installed.
     """
     import importlib.metadata