1. 概述
本文档包含了此次发布的AI应用软件包的特性，已知问题和使用限制等。
此次发布是MXC500 GPU的AI应用交付，MXMACA-Driver-3.2.1.12，MXMACA-SDK-3.2.1.10，MXMACA-Pytorch-3.2.1.3 版本适用于曦云® C500、C500-P、C500X、C550 和曦思® N260。
下表列出了系统测试覆盖率和通过率。
系统 |
测试覆盖率/通过率 |
|---|---|
maca-paddle-3.2.1.5 |
Full regression with release quality |
maca-deepspeed-3.2.1.5 |
Full regression with release quality |
maca-internlm-3.2.1.5 |
Full regression with release quality |
maca-bitsandbytes-3.2.1.5 |
Full regression with release quality |
maca-tensorflow-3.2.1.5 |
Full regression with release quality |
maca-modelzoo.cnn.training-3.2.1.5 |
Full regression with release quality |
maca-mmcv-3.2.1.5 |
Full regression with release quality |
maca-megatron-lm-3.2.1.5 |
Full regression with release quality |
maca-colossalai-3.2.1.5 |
Full regression with release quality |
maca-verl-3.2.1.6 |
Full regression with release quality |
maca-triton-inference-server-3.2.1.5 |
Full regression with release quality |
maca-modelzoo.llm.transformers-3.2.1.5 |
Full regression with release quality |
maca-diffusers.training-3.2.1.5 |
Full regression with release quality |
maca-xtuner-3.2.1.5 |
Full regression with release quality |
maca-cv-cuda-3.2.1.5 |
Full regression with release quality |
maca-vllm-3.2.1.7 |
Full regression with release quality |
maca-modelzoo.llm.vllm-3.2.1.8 |
Full regression with release quality |
maca-Ray-Patch-3.2.1.5 |
Full regression with release quality |
maca-sglang-3.2.1.8 |
Full regression with release quality |
maca-modelzoo.llm.sglang-3.2.1.8 |
Full regression with release quality |
maca-mx-swift-3.2.1.6 |
Full regression with release quality |
1.1. 交付内容
此次发布包含以下内容：
二进制文件，动态库和容器镜像文件
vLLM大模型推理框架
DeepSpeed大模型训练框架
PaddlePaddle深度学习框架
ColossalAI大模型训练框架
Megatron-LM大模型训练框架
VeRL大模型训练框架
InternLM大模型训练框架
Triton Inference Server推理框架
BitsAndBytes量化工具
TensorFlow机器学习框架
mmcv computer visios library
CV-CUDA计算机视觉加速库
XTuner大模型微调工具库
Ray开源的统一计算框架，旨在简化 Python 和 AI 应用的分布式计算
SGLang大模型推理框架
ms-swift大模型微调部署框架
《曦云® 系列通用GPU AI推理用户手册》
《曦云® 系列通用GPU AI训练用户手册》