Security Statement

System Security Hardening

1. You are advised to enable ASLR at level 2 on the system. It is also known as full randomization of address space layout. You can configure it as follows.

   echo 2 > /proc/sys/kernel/randomize_va_space
   

Recommended Runtime User

For security and least-privilege reasons, you are advised not to use root or other administrator accounts to run MindSpeed LLM.

File Permission Control

1. You are advised to set the runtime system umask value to 0027 or higher on the host and in containers. This ensures that newly created directories have a maximum default permission of 750 and newly created files have a maximum default permission of 640.
2. You are advised to apply permission controls to sensitive content such as personal data, business assets, source files, and various files saved during training. Relevant scenarios include permission control for the MindSpeed LLM installation directory and permission control for shared datasets used by multiple users. You can refer to Table 1 for the recommended permission settings.
3. MindSpeed LLM generates training data during data preprocessing and generates weight files during training. The default file permission is 640. You can tighten the permissions on generated files according to actual needs.
4. When you train with a non-root user, the permissions on your CKPT_SAVE_DIR path may be too restrictive, which prevents access to that directory. You can add chmod -R 660 $CKPT_SAVE_DIR to the script to modify the access permissions on that directory and ensure that the model weight files can be read and written normally.

Table 1 Recommended maximum permission limits for files and folders in different scenarios

Type	Linux permission reference maximum
User home directory	750 (rwxr-x---)
Program files, including scripts and library files	550 (r-xr-x---)
Program file directory	550 (r-xr-x---)
Configuration file	640 (rw-r-----)
Configuration file directory	750 (rwxr-x---)
Log files, after they complete or are archived	440 (r--r-----)
Log files, while they are being written	640 (rw-r-----)
Log file record directory	750 (rwxr-x---)
Debug files	640 (rw-r-----)
Debug file directory	750 (rwxr-x---)
Temporary file directory	750 (rwxr-x---)
Maintenance and upgrade file directory	770 (rwxrwx---)
Service data files	640 (rw-r-----)
Service data file directory	750 (rwxr-x---)
Directory for key components, private keys, certificates, and ciphertext files	700 (rwx------)
Key components, private keys, certificates, and encrypted ciphertext	600 (rw-------)
Encryption and decryption interfaces and scripts	500 (r-x------)

Data Security Statement

1. MindSpeed LLM stores model files in the checkpointing module of Megatron, and some model files use the potentially risky pickle module, which may pose data security risks.
2. During program execution, nltk.download loads corpora from the path specified by the user. Therefore, ensure network security and that the source of the downloaded package is trusted.

Runtime Security Statement

1. You are advised to write training scripts that match the available resources. If the training script does not match the resource conditions, for example, if the dataset loading memory exceeds the available memory or if the training script generates more data locally than the available disk space, errors may occur and the process may exit unexpectedly.
2. MindSpeed LLM uses PyTorch internally, and version mismatches may cause runtime errors. For details, see the PyTorch Security Statement.
3. This software uses torch.load from PyTorch to load models, and the code uses this interface with weights_only=True. For PyTorch versions <= 2.5.1, a deserialization vulnerability, CVE-2025-32434, exists. Please ensure the safety of the loaded weights and avoid malicious model loading that could attack the execution machine or device.

Public Internet Address Statement

Type	Open-source code URL	File name	Public IP address/public URL/domain/email address	Description
In-house	Not involved	mindspeed_llm/model/language_model.py:85	https://github.com/kingoflolz/mesh-transformer-jax/	Details URL
In-house	Involved	tests/test_tools/dist_test.py:6	https://github.com/microsoft/DeepSpeed/blob/master/tests/unit/common.py	Source code URL
In-house	Involved	tests/pipeline/conftest.py:6	https://github.com/microsoft/DeepSpeed/blob/master/tests/conftest.py	Source code URL
In-house	Involved	tests/ut/conftest.py:6	https://github.com/microsoft/DeepSpeed/blob/master/tests/conftest.py	Source code URL
In-house	Not involved	examples/mcore/gemma/data_convert_gemma_pretrain.sh:5	https://huggingface.co/datasets/pleisto/wikipedia-cn-20230720-filtered/resolve/main/wikipedia-cn-20230720-filtered.json?download=true	Data download URL
In-house	Not involved	mindspeed_llm/core/transformer/moe/moe_utils.py:135	https://arxiv.org/abs/2101.03961	Paper URL
In-house	Involved	mindspeed_llm/tasks/data/collator.py:4	https://github.com/OpenAccess-AI-Collective/axolotl/blob/main/src/axolotl/monkeypatch/utils.py	Source code URL
In-house	Involved	mindspeed_llm/core/distributed/distributed_data_parallel.py:126	https://github.com/NVIDIA/TransformerEngine/pull/719	Source code URL
In-house	Not involved	mindspeed_llm/core/datasets/gpt_dataset.py:159, 219	https://gitcode.com/ascend/MindSpeed-LLM/wiki/megatron data helpers可能引入的问题	Details URL

Public Interface Statement

MindSpeed LLM has not yet released a wheel package. Therefore, it does not provide any formal public interface. All functionality is invoked through shell scripts. The five entry scripts are pretrain_gpt.py, inference.py, evaluation.py, preprocess_data.py, and convert_ckpt.py.

Communication Security Hardening

Communication security hardening instructions

Communication Matrix

Communication matrix instructions

Special Scenarios

Scenario	Usage method	Port	Possible risk
When you use MindSpeed LLM for training tasks in the Megatron backend, each time it initializes the model-parallel group, it adds (3 * NPU count) random ports by default. When multiple distributed optimizers are enabled, it adds (number of distributed optimizers * NPU count) more random ports. At the same time, it configures one master-port, which is consistent with the master-port of TorchNPU.	MindSpeed LLM calls Megatron's native mpu.initialize_model_parallel function to initialize the model-parallel group and uses PyTorch distributed training APIs to start any task.	Within [1024, 65520]	Incorrect network configuration may cause port conflicts or connection issues and affect training efficiency.
The user downloads corpora through nltk.download.	The user uses nltk.download inside the code to download corpora.	Random port	If the file source is untrusted, a deserialization vulnerability may exist when the file is loaded, which can lead to file tampering.